fix(delta-sync): surface failed engine apply results instead of silently discarding them

[tanishkasinghhh]

Summary

_finalize_sync() in update_weight_from_distributed_delta.py calls
ray.get(object_refs) but discards the (success: bool, msg: str) tuples
returned by each SGLang receiver engine (via update_weights_from_disk /
update_weights_from_distributed with load_format="delta").

A failed apply — from a checksum mismatch, I/O error, or decode error — is
therefore completely silent.

Why this matters

The sender advances its pinned-CPU snapshot before flushing to the
receiver (update_snapshot_async runs ahead of the bucket flush in
_enqueue_chunk). A silently-failed apply therefore leaves the sender
snapshot permanently ahead of what the receiver actually holds. All
subsequent diffs are taken against the (already-advanced) snapshot, so the
missed weight updates are never re-sent. The drift persists until a
full broadcast reseed (process restart).

The SGLang receiver already catches all exceptions internally and returns
(False, error_msg) — the only gap was that slime never checked the return
value.

Fix

Extracts check_apply_results() into a lightweight, dependency-free module
(_apply_result_check.py) and calls it immediately after ray.get():

# Before
ray.get(object_refs)             # (False, msg) silently dropped

# After
results = ray.get(object_refs)
check_apply_results(results)     # logs per-engine errors + raises

Separating the helper into its own module keeps it unit-testable without
torch / ray / megatron / GPU.

Changes

File	What
slime/backends/megatron_utils/update_weight/_apply_result_check.py	New: the check_apply_results() helper with full docstring
slime/backends/megatron_utils/update_weight/update_weight_from_distributed_delta.py	Import + call the helper after ray.get()
tests/test_delta_sync_check_apply.py	6 CPU unit tests (no GPU/Ray needed)

Tests

python3 -m pytest tests/test_delta_sync_check_apply.py -v
# 6 passed

Closes #2104

Prepared with AI assistance (Claude Code). Every change reviewed and understood line-by-line.