[Fix] Reduce routed-expert trace memory with uint16 storage

[matrix72c]

Summary

Reduce learner memory used by rollout routed-expert traces.

• Keep routed-expert IDs resident as torch.uint16 when n_routed_experts <= 65536.
• Fall back to the existing torch.long storage path for models with more experts.
• Convert only the current layer's route slice to torch.long on the target device before router replay.

Problem

LMDeploy, the SessionServer, and the Ray object store already carry routed-expert IDs as numpy.uint16. TrainingWorker._add_rollout_routed_experts() converted the real routes to torch.long, while padding tensors also defaulted to torch.long.

This expanded every expert ID from 2 bytes to 8 bytes. A learner that prepares many long rollout samples therefore keeps a four-times-larger route payload in each training worker. Offloading changes whether that payload consumes host or device memory, but does not reduce its resident size.

Implementation

Compact learner storage

• Select torch.uint16 storage for models with at most 65536 routed experts.
• Use the selected dtype for real routes, per-sample padding, and full padding batches so that torch.cat cannot promote the result back to torch.long.
• Preserve the normal numpy.uint16 path without an additional value scan.
• Validate wider integer inputs before narrowing and reject values outside 0..65535 instead of silently wrapping.
• Preserve the original torch.long behavior when n_routed_experts > 65536.

Explicit storage/compute boundary

The MoE decoder still selects only the current layer's [local_seq_len, topk] route slice. It then converts that slice to torch.long on hidden_states.device before entering the router. The existing contiguous copy for cross-device offload is preserved.

Keeping this conversion outside the offload condition also makes the non-offload path safe: a full uint16 route tensor may be moved with the sequence context, but unsigned indices never reach gather directly.

Impact

• Route storage is reduced by 75% for models with up to 65536 routed experts.
• The only narrowing-to-index conversion is on the layer-local slice consumed by the router.
• No configuration or public API is added.
• LMDeploy, SessionServer, Ray object-store, batching, and trajectory-logging behavior are unchanged.
• Models with more than 65536 routed experts retain the previous torch.long path.

This is the narrow-dtype mitigation discussed in #2025. It does not change batch materialization or sequence-parallel transfer ordering, so those broader improvements remain separate work.

Tests

• pytest tests/rl/test_routed_experts_dtype.py tests/rl/test_prepare_train_data.py -q: 20 passed.
• Covered dtype selection at 256, 257, 65536, and 65537 experts.
• Covered real Ray-ref input, padding-only input, and mixed concatenation.
• Covered exact preservation of IDs 0, 255, 256, and 65535.
• Covered rejection of negative and overflowing wider integer inputs.
• Covered offload and non-offload decoder paths, including non-contiguous layer slices and use as gather indices.
• Ruff and git diff --check: passed.

Related to #2025.