[xLSTM] Fix train mode crashes with return_last_states=False and use_cache

[Heatdh]

What does this PR do?

Fixes #47013

In train mode with return_last_states=False, xLSTMBackend.forward returned a bare tensor while its only caller, xLSTMLayer.forward, always unpacks h, state = .... The forward pass crashed for every batch size (batch size 2 mis-unpacked the hidden states along the batch dimension and tripped the shape assertion; every other batch size failed to unpack).

This is an alternative to #47026 (thanks @lcheng321 for the first patch!). While writing regression tests for the tuple normalization, two follow-on crashes surfaced on the same path, which this PR also fixes:

1. train_with_padding + default use_cache=True: the padded kernels cannot produce last states, so there is nothing to fill the cache with — with the tuple normalization alone, xLSTMModel.forward fails at the cache copy with TypeError: 'NoneType' object is not subscriptable. Now use_cache falls back to False with a warning_once, and the cache-copy loop guards against rnn_state=None.
2. train mode + default use_cache=True + backward: xLSTMModel.forward fed the cache's state tensors into the autograd graph as initial states and then mutated them in place via copy_, so loss.backward() failed with RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation. An all-zero initial cache state is equivalent to no state, so it is no longer passed into the layers, and states are detached before being stored in the cache.

Changes

• xLSTMBackend.forward (native path) always returns an (h, last_states) tuple, matching the inference branch. The chunkwise train kernels compute the last states as a byproduct either way, so they are always requested; train_with_padding returns (h, None) since padding pollutes the last state. Return type annotation and docstring updated accordingly.
• xLSTMModel.forward disables use_cache under train_with_padding (with a warning) and skips the initial all-zero cache state so the in-place cache update cannot break the backward pass.
• xLSTMCache.__init__/reset initialize rnn_state_initial = True — this attribute was previously assigned in xLSTMModel.forward but never initialized, so it did not exist on fresh caches.
• Regression tests: train-mode forward + backward with return_last_states=False for batch sizes 2 and 3, and a train_with_padding forward with default settings.

Like #47026, this keeps the fix local to the xLSTM files; routing the last states through the output recorder (as suggested in #47013) can be a follow-up.

Testing

pytest tests/models/xlstm/test_modeling_xlstm.py -k "train_mode_without_return_last_states or train_with_padding_mode_forward"
# on main: 2 failed
# with this PR: 2 passed

pytest tests/models/xlstm/test_modeling_xlstm.py
# 90 passed, 177 skipped, 737 subtests passed
# (test_can_load_with_device_context_manager and test_can_load_with_global_device_set fail
#  identically on unpatched main in my environment — unrelated to this change)

Before submitting

• This PR fixes a typo or improves the docs (if applicable)
• Did you read the contributor guideline?
• Was this discussed/approved via a Github issue? [xLSTM] Crash in train mode with return_last_states=False: backend returns bare tensor but xLSTMLayer always unpacks (h, state) #47013
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@vasqu @ArthurZucker

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: xlstm

[Heatdh]

Thanks @lcheng321 for the initial efforts, i rerun my training and i discovered some issues along the way with the patch. @vasqu I added 2 additional tests and inspected manually the regression results and it is stable. The manual code is auto refactored to match the implementation guidelines

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 28674037288
Result: success | Grafana metrics are not available yet.