fix(fsdp): apply pause mode during non-colocated weight sync

[shangrilar]

Summary

• forward pause_generation_mode to every rollout engine in the FSDP weight-update lifecycle
• wait for all engines to pause before conditionally flushing their caches
• skip the cache flush for in_place updates
• cover abort, retract, and in_place with a parameterized lifecycle regression test

Problem

The FSDP updater called pause_generation() without forwarding the configured mode and always called
flush_cache(). Because the SGLang wrapper defaults to retract, a fully async FSDP run configured with
--pause-generation-mode in_place behaved like retract + flush instead of preserving its in-flight
requests and KV cache.

The updated lifecycle is:

abort/retract: pause(mode) -> flush -> begin weight update
in_place:      pause(mode) -> begin weight update

The separate waits also ensure every rollout engine has completed its pause before a cache flush or
weight-update session begins.

Scope

Fully async training rejects colocation, so FSDP uses the non-colocated distributed weight-sync path fixed
here. This change does not remove the colocated tensor updater's separate per-bucket cache flush and does
not claim universal colocated in_place support.

PR #1861 rewrites the same FSDP lifecycle as part of a larger SGLang client refactor, but currently retains
the no-argument pause and unconditional flush behavior.

Validation

• focused tests/fast/backends/test_fsdp_update_weight.py: 5 passed in a local Python 3.12 harness that
  loads the production updater and test file while stubbing unavailable Ray/SGLang imports
• pre-commit run --files miles/backends/fsdp_utils/update_weight_utils.py tests/fast/backends/test_fsdp_update_weight.py
• git diff --check

The full Miles dependency suite is left to upstream CI.

[shangrilar]

Ready for review. The fork workflows are currently awaiting approval. Could a maintainer apply a run-ci label when convenient? Thank you!