What does this PR do?

Decouple IS correction granularity from the GSPO policy ratio so they can be
configured independently.

Currently sequence_level_importance_ratios=True bundles two behaviors into a
single flag:

1. GSPO policy ratio: exp(mean(log(π_curr/π_prev))) — numerically stable
2. IS correction weight: exp(sum(prev - gen)) — collapses for long
   sequences

For long sequences (~9k+ tokens), even a small per-token logprob mismatch between
training and inference backends (e.g., ~2% with vLLM + Megatron) accumulates via
exp(sum) to near-zero IS weights, effectively reducing the learning rate by
10–50x with high variance across steps.

This PR adds a new is_correction_level config field to ClippedPGLossConfig
that controls IS correction weight computation independently of the PPO ratio
(which remains controlled by sequence_level_importance_ratios).

Supported values:

• "sequence" — exp(sum(prev - gen)), one weight per sequence (current
  behavior)
• "sequence_mean" — exp(mean(prev - gen)), geometric mean, numerically stable
  for long sequences
• "token" — exp(prev - gen) per token, token-level correction

When unset (None), behavior is derived from sequence_level_importance_ratios
for full backward compatibility.

Issues

None — discovered during GRPO training with Qwen3.5-35B-A3B (MoE, ~9k avg
response tokens) where sampling_importance_ratio showed mean ~0.4, std ~0.27,
with some steps collapsing to 0.02.

Usage

loss_fn:
  # GSPO policy ratio (unchanged)                                                
  sequence_level_importance_ratios: true
  token_level_loss: false                                                        
  # IS correction now independently configurable                               
  is_correction_level: "sequence_mean"  # or "token", or "sequence" (default)    
  use_importance_sampling_correction: true                                       
  truncated_importance_sampling_ratio: 2                                         
  truncated_importance_sampling_type: tis                                        

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed /NVIDIA-NeMo/RL/blob/main/CONTRIBUTING.md
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our
  /NVIDIA-NeMo/RL/blob/main/docs/testing.md for how to run tests
• Did you add or update any necessary documentation? Visit our
  /NVIDIA-NeMo/RL/blob/main/docs/documentation.md for how to write, build and test
  the docs.

Additional Information

• All 44 existing tests pass (backward compat verified)
• New test test_clipped_pg_loss_gspo_is_correction_level_sequence_mean validates
  that exp(mean) produces IS weights closer to 1.0 than exp(sum) and that the loss
  value matches hand-calculated expectations
• seq-mask-tis compatibility assertion updated to account for the new field