Fix (scaling)!: clamp to avoid inf/nan in forward/backward

[Giuseppe5]

Reason for this PR

Restrict scaling with log op could cause inf/nan in the forward pass and/or in the gradients.

Changes Made in this PR

Applying a restriction before (new behaviour) and after (existing behaviour) helps avoiding that.

Testing Summary

Add tests for standalone and runtime scaling to check inf/nan in forward/backwards.

Risk Highlight

This is a breaking change. Old checkpoints will work, new QAT/PTQ runs could produce different results from before.

• This PR includes code from another work (please detail).
• This PR contains API-breaking changes.
• This PR depends on work in another PR (please provide links/details).
• This PR introduces new dependencies (please detail).
• There are coverage gaps not covered by tests.
• Documentation updates required in subsequent PR.

Checklist

• Code comments added to any hard-to-understand areas, if applicable.
• Changes generate no new warnings.
• Updated any relevant tests, if applicable.
• No conflicts with destination dev branch.
• I reviewed my own code changes.
• Initial CI/CD passing.
• 1+ reviews given, and any review issues addressed and approved.
• Post-review full CI/CD passing.

[Giuseppe5]

I believe self.clamp_scaling is not necessary here because we're already in log domain, and self.restrict_scaling is responsible only for rounding and potential conversion from log to real domain (i.e., exponential).

Applying clamping before restrict_val means we can't get to negative values, thus scaling values between 0 and 1.
@nickfraser

[nickfraser]

Sounds right to me!