feat: HYBRID_SHARD draft training across colocated target islands

[maocheng23]

Motivation

Stack 5/5 replacing draft #766. K3-class colocated runs put one TP8 target island on each node; sharding the draft across the whole world would make every FSDP parameter all-gather cross nodes. HYBRID_SHARD uses the target-TP group as the FSDP shard group and the target-DP group for replication, keeping parameter traffic node-local while only replica synchronization crosses nodes.

Modifications

• training.fsdp_sharding accepts HYBRID_SHARD, valid only for online local_colocated with training.tp_size > 1 and at least two target islands.
• ParallelConfig carries explicit reduction_process_group (WORLD — the HSDP fsdp_process_group is a (shard, replica) pair and cannot serve as a collective group) and grad_norm_process_group (the shard group, counting each replicated parameter shard exactly once). TrainerCore reduces losses/metrics through a single _reduction_group() helper with the FSDP-group fallback for every non-HSDP topology — feat: add scalable colocated online training #766's __dict__-based mock-dodging lookup is replaced by fixing the test's mock contract instead.
• Adds the Kimi K3 4x8 B300 TP8/HSDP starting recipe and the HSDP-across-islands documentation section.

Related Issues

Splits #766. Stack: #1 capture-rows ← #2 teardown-abort ← #3 rank0-tracker ← #4 colocated-core ← #5 (this).

Accuracy Test

• HSDP is configuration- and process-group-tested (ParallelConfig.from_distributed group wiring, grad-norm group selection, schema validation); the intended 4x8 B300 K3 run has not been executed yet — this PR stays draft until it has, matching feat: add scalable colocated online training #766's open checklist item.

Checklist

• Format your code according to the Code Formatting with Pre-Commit (black --check and isort --check-only).
• Add unit tests.
• Update documentation and example recipes.
• Validate Kimi K3 with TP8/HSDP on 4x8 B300 before marking ready for review.