refactor: rename residual DiffRL identifiers to UniRL

[CjhHa1]

Summary

• Breaking rename of leftover DiffRL identifiers onto UniRL: DIFFRL_* env vars, socket/lock/shm paths, LoRA constants, worker RPC names, mixin helpers, and patch sentinels now use UNIRL_* / _unirl_*.
• No old-name fallback. DIFFRL_DISABLE_CUDNN is removed; only the existing UNIRL_DISABLE_CUDNN remains. LoRA adapter id stays 1, serialize defaults stay on, replica rank still defaults to 0.
• Cleanup in the touched weight-sync code: drops the unused _unirl_describe_params / _unirl_param_checksums worker RPCs, redundant int() around the int LoRA id constant and the checksum adapter_id, the unreachable str(adapter_name) or UNIRL_LORA_NAME fallback, the NCCL mixin's load_weights fallback (every extension also mixes in BucketedIPCReceiveMixin), a pointless import alias, inlines the single-use parameter-module lookup into _unirl_loaded_param_checksums, and registers the vLLM-Omni LoRA under UNIRL_LORA_NAME (matching the generation request) so the backend no longer takes adapter_name.

Related Issue

N/A. Follows #413 (merged); this branch is rebased onto current main (fbbf829). The rebase conflict was in patches/runtime.py: #474 deleted patch_per_request_ar_seed, so the rename inside it was dropped.

Test Plan

• rg -n -i 'diffrl|diff_rl|diff-rl' is empty on the rebased tree
• Every renamed worker RPC name matches its collective_rpc / _call_rpc call site
• python3 -m compileall on the changed files
• SKIP=no-commit-to-branch pre-commit run --from-ref <base> --to-ref HEAD passes
• NCCL receive path with a stubbed dist.broadcast loads through _unirl_load_weights
• _unirl_loaded_param_checksums and _unirl_loaded_lora_checksums on stub workers, run against the pre-PR implementation (git show <base>:...) on the same inputs: identical results for direct / empty-wrapper / no-module workers, names hit / miss, top-level and model_runner LoRA managers, and an unknown adapter id
• Not run: GPU rollout smoke; reason: pure rename plus removal of unreachable/unused code, no behavior change on live paths

Compatibility / Risk

• Breaking for launch scripts: any DIFFRL_* export (DIFFRL_IPC_DIR, DIFFRL_REPLICA_RANK, DIFFRL_MODEL_LOAD_SERIALIZE, DIFFRL_OMNI_BOOT_SERIALIZE, DIFFRL_DISABLE_CUDNN) must switch to the UNIRL_* name; old names are silently ignored.
• Socket, lock, and shm names change (unirl-zmq-*.sock, /tmp/unirl_*.lock, unirl_weights_*); trainer and rollout must run the same revision.

Reviewer Notes

The first commit is the mechanical rename; the follow-up commits are the cleanup and can be reviewed separately.

Pre-push review (independent reviewer: GPT-5.6 Sol, no findings):

• Fixed: int(adapter_id) was applied three times along the LoRA-checksum path (WeightSync, backend, worker RPC); every caller passes the int UNIRL_LORA_INT_ID.
• Kept: adapter_name on WeightSync.set_lora_from_tensors* (it is the RolloutEngine interface, also used by SGLang, and names the adapter in the wake-reload log/error); getattr(..., "_unirl_*", False) idempotency sentinels on monkey-patched callables.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.