Enable shared memory prefetch for DCP weight sync

[casteryh]

Summary

This PR enables shared memory multiprocess prefetch for DCP-based weight synchronization, resolving the TODO at generator.py:442.

Changes

Core functionality:

• Added _fetch_weights_parallel() method - Unified parallel fetching logic for both DCP and non-DCP paths
• Extended _WeightFetcher.fetch() method - Parameterized to handle both torchstore individual tensors and DCP checkpoints
• Updated Generator.update_weights() - Automatically detects DCP usage and uses appropriate shared memory prefetch

Refactoring:

• Reduced ~80 lines of code duplication by consolidating common parallelization and SharedTensor wrapping logic
• Renamed checkpoint_id parameter to dcp_key for clarity (it's the torchstore key, not the filesystem checkpoint path)
• Each fetcher now calls ts.get(dcp_key) to retrieve the DCP handle with metadata

Bug fixes:

• Removed premature dcp_handle.drop() call - checkpoint files must be preserved for recovery

Performance Benefits

This change enables the same performance benefits for DCP-based weight sync that were previously only available for individual tensor loading:

• Parallel loading across 8 fetcher processes (configurable via n_fetcher_procs)
• Shared memory - single copy instead of N worker copies
• Reduced latency during weight updates

Technical Details

DCP path:

1. Generator calls ts.get(dcp_whole_state_dict_key) to get param names
2. Passes the torchstore key to fetchers for parallel loading
3. Each fetcher calls ts.get(dcp_key) to get the DCP handle (metadata + checkpoint path)
4. Tensors loaded via load_tensor_from_dcp() and wrapped in SharedTensor
5. Workers load from shared memory instead of each calling DCP load

Non-DCP path:

1. Generator queries torchstore keys to get param names
2. Passes version number to fetchers for parallel loading
3. Each fetcher loads individual tensors via ts.get(param_key)
4. Tensors wrapped in SharedTensor and handles returned
5. Workers load from shared memory instead of each calling ts.get()

Testing

• Syntax validation passed
• All pre-commit hooks passed (trailing whitespace, AST check, flake8, ufmt, pydoclint)
• No breaking changes to existing API

---

This PR was created by Claude Code on behalf of @casteryh