integration: add TransformersProcessRewardModel with RoPE buffer repair

[lambdabaa]

What this adds

TransformersProcessRewardModel — a new integration class for
classifier-head PRMs (Qwen2.5-Math-PRM-7B and compatible models) using
standard transformers + torch, with no dependency on a running vLLM
server.

Before this PR, its_hub's only in-process PRM integration was
LocalVllmProcessRewardModel, which requires a running vLLM server.
This adds a second path: any machine with transformers and torch
installed (MPS, CUDA, or CPU) can now load and score trajectories with
a classifier-head PRM in a single in-process call.

Bug fixed: RoPE buffers zeroed by transformers ≥5 meta-tensor init

Without the repair, loading Qwen2.5-Math-PRM-7B via
AutoModel.from_pretrained(..., trust_remote_code=True) with transformers
≥5 produces a constant score of ~0.50003 for every step regardless of
input — the model appears to work but all scores are meaningless.

Root cause: transformers 5.x uses meta-tensor initialisation during
from_pretrained. Non-persistent buffers (inv_freq, cos_cached,
sin_cached) on Qwen2RotaryEmbedding are materialised as zeros rather
than being computed, and load_state_dict skips them because they are
non-persistent. The result is an identity-collapsed rotary embedding that
projects every token to roughly the same position.

_repair_rotary_embeddings() detects and fixes this post-load by
recomputing inv_freq from module.base and module.dim for every
RotaryEmbedding submodule. No other changes to model weights are made.

Other implementation notes

• Trailing <extra_0> separator: "<extra_0>".join(steps) + "<extra_0>"
  gives exactly N separator tokens for N steps, matching the model's training
  format.
• bfloat16 default: avoids NaN hidden states that fp16 produces on MPS
  for Qwen2ForProcessRewardModel's deep backbone.
• use_cache=False: skips DynamicCache.from_legacy_cache churn in
  transformers 5.x for inference-only scoring.

Tests

Six tests added covering:

• ImportError guard when torch/transformers are absent
• _repair_rotary_embeddings() recomputes zeroed inv_freq buffers
• score() returns len(steps) floats in [0, 1]
• Trailing <extra_0> separator is present (N steps → N separators)
• Class is importable from its_hub.integration

[beatsmonster]

@lambdabaa could we please rebase to main/v1