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

[JamesKunstle]

Adds requirements-hpu.txt file and optional-dependencies.hpu to pyproject.toml

Sets lower bound for accelerate for rocm, cuda, and hpu separately.

For HPU, sets accelerate>=0.33.0 because optimum-habana requires this.

For CUDA/ROCm sets accelerate=>0.34.2 because accelerate got FSDP patches that we need.

[JamesKunstle]

Blocked because we probably need handling for DeepSpeed not being installed for ROCm and Gaudi: #291

[JamesKunstle]

Requires instructlab/instructlab#2641

[Maxusmusti]

needs to be deleted

[Maxusmusti]

the general overrides the extras requirements, so when this exists:
pip install instructlab-training: Requirement already satisfied: accelerate
pip install instructlab-training[cuda]: Requirement already satisfied: accelerate

when removed:
pip install instructlab-training: N/A
pip install instructlab-training[cuda]: Requirement already satisfied: accelerate>=0.34.2

[Maxusmusti]

trl also has accelerate as a dependency as well, so just having trl>=0.9.4 in the requirements.txt makes sure general accelerate is pulled in.

From trl==0.9.4 requirements.txt:

datasets>=1.17.0
torch>=1.4.0
tqdm
transformers
accelerate
peft>=0.3.0
tyro>=0.5.7

[Maxusmusti]

Requirement already satisfied: accelerate in .../lib64/python3.11/site-packages (from trl>=0.9.4->instructlab-training==0.6.0a1.dev7) (0.33.0)

[RobotSail]

LGTM