Phi‑2 support: partial RoPE, deterministic dropout, loader dispatch, and smoke test

[MagellaX]

PR description

Adds first-class support for Microsoft Phi-2 with robust, minimal integration.

What/why

• Implement FastPhiModel for Phi-2; enable Unsloth fastpaths.
• Handle partial RoPE correctly (applies rotation to first rotary_dim features, default 0.4 when absent).
• Provide deterministic, stateless residual dropout (device-agnostic, seedable) for attention/MLP outputs.
• Wire loader dispatch and add alias mapping for 4-bit loading.

Highlights

• unsloth/models/phi.py: Phi attention forward (partial RoPE), CausalLM fastpath, post-patch defaults, deterministic dropout attachment.
• unsloth/models/loader.py: dispatch model_type == "phi"; call model post_patch.
• unsloth/models/mapper.py: add unsloth/Phi-2-bnb-4bit alias.
• unsloth/kernels: new deterministic dropout; partial RoPE helpers; safe LayerNorm/GeLU hooks (torch-backed by default).
• tests/qlora/test_unsloth_qlora_train_and_merge.py: Phi-2 smoke test (loads and runs forward; skips if weights unavailable).

Compatibility

• No behavior changes to non-Phi models.
• Kernel hooks are optional and torch-backed by default.

Links

• Addresses Phi-2 finetuning request: Issue Any plans to support phi-2 finetuning? #85
• Related prior effort (not wholesale imported): PR Staging PR for implimenting Phi-2 support. #97

[Datta0]

Thanks for the contributions. A few minor comments
Please remove the unnecessary pycache files .

[MagellaX]

Hey @danielhanchen any thoughts bro?