fix: model.forward now accepts return_dict via kwargs

Since huggingface/transformers#36794 it handles
the return_dict with can_return_tuple decorator that expects the
argument to be passed as kwarg.

Without the patch, instructlab functional tests fail with:

AttributeError: 'bool' object has no attribute 'unsqueeze'

This is because the passed return_dict argument is incorrectly
interpreted as cache_position.

Of course, the way the model forward method is overridden here is quite
problematic and would benefit from a refactor. This patch is a stop-gap
to fix the CI with minimal changes.

Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>

provide more docstrings

Signed-off-by: Oleg Silkin <97077423+RobotSail@users.noreply.github.com>

Merge pull request #407 from fabiendupont/hpu-remove-optimum-deps

Remove optimum dependency

Merge pull request #350 from instructlab/deepspeed-move

Update Dependencies to Move DeepSpeed to CUDA Extras

Merge pull request #320 from JamesKunstle/hpu-config

deps: change hf accelerate bound, add requirements-hpu.txt

Merge pull request #324 from jaideepr97/move-tokens

chore: move token classes into chat templates

Merge pull request #312 from instructlab/granite-dolomite

Adding Dolomite Support and Bringing HF Padding-Free into Performance Parity

Merge pull request #282 from instructlab/mergify/bp/release-v0.5/pr-217

chore: add exit code & tox fix (backport #217)

Merge pull request #260 from instructlab/mergify/bp/release-v0.5/pr-259

Add rocm extra to pyproject.toml (backport #259)

Merge pull request #258 from instructlab/mergify/bp/release-v0.5/pr-255

fix: Add explicit flash_attn requirement for ROCm (backport #255)