fix: filter single-file safetensors by assigned layers before push

[cjchanh]

Problem

When a Cake master distributes a single-file safetensors model to a worker, it pushes the entire file regardless of how many layers the worker is assigned. For Qwen2.5-7B-Instruct-4bit (4 GiB single file), an iPad worker with a 3 GiB jetsam budget receives the full 4 GiB, exceeds memory, and crashes with early eof.

The indexed model path (model.safetensors.index.json present) already filters correctly via weight_map. The single-file fallback at sharding/mod.rs unconditionally adds model.safetensors to the push list.

Fix

For single-file models with assigned layers, the push path now:

1. Reads only the safetensors header to enumerate tensor names
2. Filters tensors by assigned layer prefixes (same starts_with logic as the indexed path)
3. Calls extract_layer_tensors to build a minimal safetensors blob containing only the needed tensors
4. Pushes the reduced blob instead of the full file

Backward compatible: if layers is empty (no specific assignment), the full file is still pushed. If no tensors match assigned layers, falls back to full push with a warning.

Results

Tested with M5 Max master + iPad Air M3 worker, Qwen2.5-7B-Instruct-4bit:

Metric	Before	After
Push size	4 GiB (full model)	250.1 MiB (52 tensors, 2 layers)
iPad RSS	jetsam kill	1.4 GiB (under 3 GiB limit)
Result	crash (early eof)	coherent output at 17.21 tok/s

Test plan

• cargo test -p cake-core --lib — 641 tests pass (638 existing + 3 new)
• cargo test -p cake-core --test unit — 235 tests pass
• cargo clippy — zero new warnings
• Integration: M5 master + iPad Air M3, 2 layers of 7B-4bit, verified 250.1 MiB push, 1.4 GiB RSS, correct inference
• Extended inference: longer generation to verify sustained correctness across distributed layers

New unit tests

• extract_layer_tensors_single_file_filters_correctly — 4 tensors in, request 2, verify only 2 in output with correct data bytes
• extract_layer_tensors_single_file_all_layers — request all tensors, verify all present with correct total size
• extract_layer_tensors_single_file_missing_tensor_errors — request nonexistent tensor, verify error

[cjchanh]

This fix is still relevant from my side. I attempted a conflict-only rebase against current main but found that recent upstream changes (PR #84's iOS TCP retry refactor and adjacent commits) introduce API drift beyond a simple merge — Strategy::assign_layers trait signature changed (7→8 params), Message::DeviceInfoRequest variant was removed, and the BUILD_HASH constant location shifted, producing 16 compile errors when ee01115 is rebased onto current main. Rather than ship a broken-build force-push, I'm leaving this PR in CONFLICTING state. Happy to either redo this as a fresh PR against current main (cherry-picking only the minimal safetensors-filter logic) or close this in favor of that — let me know which you'd prefer.