Fix: skip KL backward graph when coefficient is zero

[Dodojordi]

Summary

When --use-kl-loss is enabled with kl_loss_coef == 0, slime still
constructs the reference KL autograd graph and adds the following term to the
actor loss:

loss = loss + args.kl_loss_coef * kl_loss

This is not only unnecessary computation: it can turn an otherwise valid
policy loss and its gradients into NaN. Multiplying by a zero coefficient does
not make a non-finite KL value harmless, because 0 * NaN is still NaN.

This PR keeps KL metric reporting for observability while preventing the
zero-weight KL path from participating in actor backward or contaminating the
training loss.

Motivation

Users may want to keep reference KL metrics for observability without using
KL as a training objective.

The current implementation couples metric reporting with optimization. Even
when the coefficient is zero, it retains the KL computation graph and inserts
the zero-weight term into the total loss. If the KL calculation produces a
non-finite value under extreme log-probability differences, this can propagate
NaNs through actor backward and interrupt training despite KL regularization
being disabled.

It is therefore important to make zero mean metric-only, rather than
"compute the gradient and multiply it by zero."

Changes

When kl_loss_coef == 0:

• Detach the current, old, and reference log-probabilities used by the KL
  metric path.
• Compute and report kl_loss as before.
• Do not add the zero-weight KL term to the actor loss.
• Apply the same behavior to the unbiased-KL importance-ratio path.

This prevents a non-finite metric-only KL value from poisoning the actor loss
or policy gradients.

When kl_loss_coef != 0, the existing differentiable behavior is preserved.

Behavior

Configuration	Result
use_kl_loss=False	Unchanged
use_kl_loss=True, kl_loss_coef=0	KL is metric-only
use_kl_loss=True, kl_loss_coef!=0	KL contributes to actor gradients
use_unbiased_kl=True, kl_loss_coef=0	Importance ratio is also computed without gradients

Reference-model execution and kl_loss reporting remain enabled.

Validation

A focused local regression test verified that:

• KL is still reported when the coefficient is zero.
• Tensors passed through the metric-only KL path do not require gradients.
• Policy backward remains functional.
• Policy gradients are not affected by the zero-weight KL term.

Additional checks passed:

git diff --check
python -m py_compile slime/backends/megatron_utils/loss.py

Example

--use-kl-loss
--kl-loss-coef 0.0
--kl-loss-type low_var_kl