fix: bind model_index by value in MegatronEngine's param_sync_func callbacks - #4273

What does this PR do?

MegatronEngine.get_module_config builds one callback per model chunk:

config.param_sync_func = [
    lambda x: self.optimizer.finish_param_sync(model_index, x)
    for model_index in range(len(self.module))
]

The lambda captures model_index by reference, not by value - a closure binds the variable, not whatever it happened to hold when the lambda was created. By the time any of these callbacks is actually invoked (later, once overlap_param_gather and delay_param_gather training kicks in), the comprehension has already finished and model_index holds its final value - len(self.module) - 1 - for every callback in the list. Whichever model chunk's callback fires, it reports the last chunk's index to finish_param_sync, not its own.

Fixed with the standard default-argument capture (model_index=model_index), which evaluates the current value at lambda-creation time instead of call time, so each callback gets its own bound copy - this is the same list used one line above for no_sync_func/grad_sync_func, but those bind plain method references (model_chunk.no_sync) rather than a lambda closing over the loop variable, so they aren't affected by this.

Found via codehound, a static analyzer I built that flags exactly this pattern (a lambda inside a loop that gets stored rather than called immediately).

Fixes # (no issue filed - a self-contained, narrowly-scoped correctness fix)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? (not applicable - internal training-loop wiring, no public API change)
• Did you write any new necessary tests?

get_module_config() itself needs the real megatron-core package to exercise directly (get_args/get_model_config import straight from megatron.core, which isn't available in this dev/test environment), so the new test mirrors the exact list-comprehension shape in isolation with a lightweight optimizer stand-in, and asserts each callback reports its own index. Verified it fails against the pre-fix code (all three callbacks report index 2) and passes against the fix.

Who can review?

Tagging @SunMarc per the "Core parts of the library" note above.