Steer agent to HF kernels instead of pip install flash-attn

[DarshanCode2005]

Resolves #202

The agent kept trying to pip install flash-attn in jobs, which often takes
ages to compile or fails outright on the job's CUDA/torch combo. The HF
kernels library lets you pull a prebuilt flash-attn (and friends) straight
from the Hub via attn_implementation="kernels-community/flash-attn2".

• Rewrote the HARDCODED UNAVAILABLE PACKAGES bullet in system_prompt_v3.yaml
  to recommend kernels first and only fall back to pip install when no Hub
  kernel covers the need. Listed the common kernel ids (flash-attn2,
  vllm-flash-attn3, paged-attention) so the agent doesn't have to guess.
• Fixed the kernels entry in explore_hf_docs. The old description
  ("Lightweight execution environments and notebook-style workflows") was
  describing something else entirely and would have sent the agent the wrong
  way when it actually went looking.

Refs: https://huggingface.co/docs/kernels/index, https://huggingface.co/docs/trl/kernels_hub