Replace CHOOSE_IMPLEMENTATION/CHOOSE_SVE_IMPLEMENTATION macros with constexpr dispatch tables

[dor-forer]

Summary

• Replaces the C1..C64/CASES4..64/CHOOSE_IMPLEMENTATION/SVE_CASE/CHOOSE_SVE_IMPLEMENTATION macro pyramid in implementation_chooser.h with two small C++20 helpers in the new spaces/functions/residual_dispatch.h: dispatch_by_residual<Dist, Chunk>(dim, selector) and dispatch_by_sve_residual<Dist>(dim, chunk, selector). Both build a static constexpr lookup table once, at compile time, and index it with dim % Chunk — same O(1) shape as the switch it replaces, but the chunk size is stated exactly once (the helper's own template argument), so it can't drift from a separately-written modulus at the call site the way two independent literals could.
• All ~130 Choose_* function bodies across all 21 spaces/functions/*.cpp files (x86: SSE/SSE3/SSE4/AVX/AVX2 family/F16C/AVX512 family; ARM: NEON family, SVE/SVE2/SVE_BF16) converted. implementation_chooser.h/implementation_chooser_cleanup.h deleted — zero remaining references anywhere.
• Choose_*/GetDistFunc* names and signatures are unchanged, so tests/unit/test_spaces.cpp and tests/benchmark/spaces_benchmarks/bm_spaces.h needed zero changes. Every leaf SIMD kernel is untouched.
• This is R1 of a two-part refactor (board-reviewed plan in SPACES-REFACTOR-PLAN.md/SPACES-REFACTOR-LOG.md in this repo). R2 (replacing the #ifdef-nested runtime feature-check cascades in IP_space.cpp/L2_space.cpp with a declarative dispatch table) is stacked on top of this in a separate PR.

Verification

• Compile: all 15 x86 files independently recompiled standalone with their exact per-tier release flags (-Werror -Wall -std=gnu++20 -O2, e.g. -mavx512f -mavx512bw -mavx512vl -mavx512vnni for AVX512F_BW_VL_VNNI.cpp) — clean, no warnings. The 7 ARM files (NEON/SVE families) converted with the same pattern but could not be compiled on this dev box (no aarch64/SVE cross-toolchain available) — needs a real ARM build + test pass before merge.
• Unit tests: make unit_test CTEST_ARGS="-R Spaces" — 1442/1442 passed, 0 failed, on this machine's real detected x86 tiers.
• Binary size (measured on the AVX2 BF16_IP/BF16_L2 conversion): per-wrapper code drops from 528 to 19 bytes — the 32-way jump-table landing blocks disappear in favor of one indexed load into a 256-byte static constexpr table. Net code+data across a converted pair: ~48% smaller. Section-by-section breakdown confirms the whole-.o size increase some builds show is 100% DWARF debug info for the longer template/lambda names (stripped in production); real shipped .text+.rodata+.data.rel.ro shrank.
• Dispatch speed: a synthetic microbenchmark (same switch-shape vs same array-shape, random dim per call to defeat branch-predictor warmup) shows ~5.8x faster dispatch selection (15.6ns → 2.7ns/call). Real and reproducible, but this call happens exactly once per index build (components_factory.h:23), never per query — confirmed via bm_spaces.h's own structure, which calls Choose_* once in benchmark setup, outside the timed loop. So: no effect on distance-kernel query throughput (expected and confirmed — kernels are byte-identical), real effect on the one-time dispatch cost.

Test plan

• make unit_test CTEST_ARGS="-R Spaces" — 1442/1442 passed
• Standalone compile of all 15 x86 functions/*.cpp with exact release flags
• Real ARM/NEON/SVE compile + test pass (no aarch64 toolchain on this dev box)
• make check-format / CI format check

[dor-forer]

✅ `make check-format` passes clean on this branch.