fix(configs): validate inference weight_broadcast type against orchestrator

[EazyReal]

What

validate_shared_weight_broadcast checks that the trainer, orchestrator, and (optional) inference weight-broadcast transports agree. It used a chained comparison:

trainer.weight_broadcast.type != orchestrator.weight_broadcast.type != inference.weight_broadcast.type

Python evaluates this as (trainer != orchestrator) and (orchestrator != inference), so inference is never compared against trainer. When trainer and orchestrator agree but inference differs (e.g. filesystem / filesystem / nccl), the check passes and the mismatch only surfaces as a broken weight sync at runtime.

This is reachable whenever the shared [weight_broadcast] is unset (the default): auto_setup_weight_broadcast only unifies the three sub-configs when the shared field is set, otherwise it passes the user's per-component values straight to this validator.

Why / fix

Compare each component against the orchestrator independently, which transitively requires all three to match. Both original error messages are preserved.

if trainer.weight_broadcast.type != orchestrator.weight_broadcast.type:
    raise ValueError("Trainer ... and orchestrator ... are not the same ...")
if inference is not None and inference.weight_broadcast.type != orchestrator.weight_broadcast.type:
    raise ValueError("Inference ... and orchestrator ... are not the same ...")

Test

tests/unit/test_configs.py:

• test_weight_broadcast_inference_mismatch_raises — (filesystem, filesystem, nccl) now raises (fails on main: "DID NOT RAISE").
• test_weight_broadcast_all_matching_passes — all-matching config still validates.

All 111 tests in test_configs.py pass; ruff format/ruff check clean.

[EazyReal]

@mikasenghaas @samsja Could I get a quick review on this config validation fix? It catches inference weight-broadcast type mismatches before runtime weight sync breaks.