fix(trainer): clip gradients before optimizer.step() so --max-grad-norm takes effect

[valter-silva-au]

Summary

Fixes #1366.

--max-grad-norm is meant to clip gradients for training stability, but every
trainer calls th.nn.utils.clip_grad_norm_ after optimizer.step():

self.optimizer.zero_grad()
loss.backward()
rt_profiler.record('train_backward')
self.optimizer.step()          # <-- optimizer already used the un-clipped grads
rt_profiler.record('train_step')

if max_grad_norm is not None:
    th.nn.utils.clip_grad_norm_(model.parameters(), max_grad_norm, grad_norm_type)

Since optimizer.step() has already consumed the gradients to update the
parameters, clipping afterward is a no-op for that update — the gradients are
zeroed at the start of the next iteration before they're ever used. The result
is that --max-grad-norm silently does nothing.

The correct contract is backward → clip → step, so the optimizer applies the
clipped gradients.

Changes

Moved the clip block to run immediately after loss.backward() and before
optimizer.step(). The issue names np_trainer, but the same wrong ordering is
present in all five trainers, so I've fixed them together:

• python/graphstorm/trainer/np_trainer.py
• python/graphstorm/trainer/ep_trainer.py
• python/graphstorm/trainer/lp_trainer.py
• python/graphstorm/trainer/mt_trainer.py
• python/graphstorm/trainer/glem_np_trainer.py

This is a pure statement reorder — no signature or API change. The
if max_grad_norm is not None: guard is preserved, so the default
(max_grad_norm=None) path is unchanged; behavior only changes when a user
explicitly sets the flag, where it was already broken.

Scope note for maintainers: the issue only mentions np_trainer. I scoped
the fix to all five trainers since the bug is identical in each, but I'm happy
to narrow this PR to just np_trainer (or np/ep/lp) if you'd prefer a smaller
change — just let me know.

Testing

Added tests/unit-tests/test_trainer.py::test_node_trainer_grad_clip_before_step:
it spies on clip_grad_norm_ and GSOptimizer.step (both call through to the
originals so fit() runs for real), runs trainer.fit(num_epochs=1, max_grad_norm=1.0) on the existing tiny CPU/gloo dummy graph, and asserts the
clip is observed before the step on every iteration.

• Fails on the current ordering (records ['step', 'clip']).
• Passes after the reorder.
• Full test_trainer.py suite: 17 passed (CPU/gloo, world_size=1).
• pylint --rcfile=tests/lint/pylintrc python/graphstorm/trainer/: 10.00/10.