Fix RLHF _template_context clears template.max_length by default causing truncation to be skipped in GRPO/GKD training

[burm95]

Background
In GRPO/GKD training, users expect max_length + truncation_strategy=left to truncate long samples before model forward, preventing OOM.

However, entering the _template_context default path (i.e. with self._template_context(template):) unconditionally sets template.max_length = None via the default argument. Because Template._encode_truncated only performs truncation when self.max_length is not None, the truncation step is effectively skipped for the re-encoded training batch.

This leads to the user-visible symptom: even with a very small max_length and truncation_strategy=left, samples that exceed the configured length still cause OOM during policy/ref forward.

Fix
Use a sentinel object (_TEMPLATE_CONTEXT_USE_ORIGINAL) to distinguish between:

Not passed (default): preserve the original template.max_length so truncation keeps working.
Explicitly passed: use the caller-provided value, including None, to temporarily un-set the length limit.
This preserves the existing helper semantics (max_length=None can still explicitly disable length override) while fixing the regression where the default context entry silently disables truncation.

Files changed
swift/rlhf_trainers/rollout_mixin.py
Add _TEMPLATE_CONTEXT_USE_ORIGINAL sentinel.
Update _template_context signature and logic to only override max_length when the caller explicitly passes it.
Verification
Manually validated the call sites in GRPOTrainer/GKDTrainer that enter _template_context without an explicit max_length: they no longer clear template.max_length, so Template._encode_truncated correctly applies the configured max_length + truncation_strategy=left.
Local runs with long prompt samples previously causing OOM now correctly truncate and fit into memory under the same max_length setting.
Notes
This is a minimal, behavior-preserving fix for the default code path. It does not change the call signature contract for code that explicitly passes max_length=None.

[hjh0119]

max_length in GRPO/GKD bounds the prompt (already enforced on the rollout side), so re-encoding prompt+completion here must not truncate again.