perf(build): batch generated dispatch builds

[voltjia]

Summary

• Split generated dispatch implementation definitions across multiple generated source files.
• Keep one shared generated_dispatch.h declaration header while batching per-operator dispatch definitions.
• Mark the shared IndexToOffset header helper inline so generated dispatch shards can include the same implementation headers safely.
• Add INFINIOPS_DISPATCH_BATCH_SIZE to tune generated dispatch source batch size.

Motivation

The generated dispatch translation unit can become large as bindings grow. Splitting it into multiple generated source files lets the build system compile dispatch definitions in parallel and reduces single-translation-unit compiler pressure.

Closes # N/A

Type of Change

• N/A feat — new feature / new operator / new platform
• N/A fix — bug fix
• perf — performance improvement
• N/A refactor — code restructuring without behavior change
• N/A test — adding or fixing tests only
• N/A docs — documentation only
• N/A build / ci — build system or CI configuration
• N/A chore — tooling, formatting, or other non-code changes
• N/A Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Test Results on Supported Platforms

Direct profile, PYTEST_WORKERS=1.

Platform	Built	pytest Result	Notes / Hardware
NVIDIA	Yes	4151 passed, 1375 skipped in 312.73s	Full pytest tests/. Matches #604/#605/#606/#607/#608 pass/skip counts.
Iluvatar	Yes	3651 passed, 375 skipped in 259.43s	Full pytest tests/. Matches #604/#605/#606/#607/#608 pass/skip counts.
MetaX	Yes	5795 passed, 1447 skipped in 351.71s	Full pytest tests/. Matches #604/#605/#606/#607/#608 pass/skip counts.
Cambricon	Yes	3073 passed, 3857 skipped in 891.89s	Full pytest tests/. Preserves #607's fix for the previous integer-input failures.
Moore	Yes	300 failed, 5459 passed, 1483 skipped in 536.58s	Full pytest tests/. Same known pre-existing tests/test_gemm.py MUSA failures as #604/#605/#606/#607/#608.
Ascend	Yes	3828 passed, 138 skipped in 509.84s; container exit code 137 after pytest	Full pytest tests/. No pytest failures in this run; same post-test exit-code behavior as #605/#606/#607/#608 and no regression versus #604.

Compared with the last merged baseline PRs #604, #608, #605, #607, and #606, this PR has no regression in build status, collected coverage, or failure count.

Full validation summaries

ruff:
1 file already formatted
All checks passed!

clang-format:
clang-format --dry-run --Werror src/common/generic_utils.h

generation smoke:
generated/bindings/generated_dispatch_0.cc
generated/bindings/generated_dispatch_1.cc

nvidia build=0 test=0
4151 passed, 1375 skipped in 312.73s

iluvatar build=0 test=0
3651 passed, 375 skipped in 259.43s

metax build=0 test=0
5795 passed, 1447 skipped in 351.71s

cambricon build=0 test=0
3073 passed, 3857 skipped in 891.89s

moore build=0 test=1
300 failed, 5459 passed, 1483 skipped in 536.58s

ascend build=0 test=137
3828 passed, 138 skipped in 509.84s

Benchmark / Performance Impact

Expected to improve build parallelism for generated dispatch bindings. Runtime behavior is unchanged.

Notes for Reviewers

This only changes generated binding source layout. generated_dispatch.h still declares the same entrypoints, while generated source files are split by batches of operators. IndexToOffset is made inline because generated dispatch shards can now include the same implementation headers in multiple translation units.

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz.
• Each commit message follows Conventional Commits.
• Small PR is a single squashable commit.
• No stray merge commits from master.
• No fixup! / squash! / wip commits remain.

Scope and Design

• Changes are minimal.
• No dead code, debug prints, or unrelated TODOs were added.
• No unrelated formatting churn.
• N/A Public API changes.

General Code Hygiene

• Code is self-explanatory.
• Modified files end with a newline.
• No trailing whitespace, tab/space mixing, or stray BOMs.
• Identifiers in comments are wrapped in backticks.
• Comments are in English.
• Comments are complete sentences.

C++ Specific

• Code follows the existing C++ style.
• clang-format was run on the modified C++ header.
• N/A clang-tidy was not run for this small helper linkage fix.
• N/A Operator parameter order.
• No exceptions are thrown.
• N/A Error and warning messages.
• N/A Kernel file naming.
• N/A Kernel/launcher separation.
• N/A Constructor initializer list order.
• Member/namespace spacing follows surrounding style.
• N/A New operators.
• N/A Raw new/delete.

Python Specific

• Code follows the surrounding Python style.
• ruff check scripts/generate_wrappers.py passed.
• ruff format --check scripts/generate_wrappers.py passed.
• N/A New comments.
• No blank line rule violations were introduced.
• N/A Docstrings.
• N/A Type hints.

Testing

• Full platform testing was run across NVIDIA, Iluvatar, MetaX, Cambricon, Moore, and Ascend.
• N/A No new tests were added.
• N/A No test parameterization changes.
• N/A No new flaky tests.
• Regression coverage: full-platform builds compile and link the generated dispatch shards.

Build, CI, and Tooling

• Fresh package build passed on all supported platforms.
• N/A compile_commands.json.
• N/A Backend auto-detection.
• N/A CUDA-like mutual exclusion.
• N/A CI workflows are expected to validate formatting.
• N/A Runtime dependencies.

Documentation

• N/A User-facing documentation changes.
• N/A New public utilities.
• N/A Breaking changes.

Security and Safety

• No secrets or personal data are committed.
• N/A Third-party code.
• N/A Unsafe pointer arithmetic.

[voltjia]

请 @wooway777 初审，@Ziminli 终审。