Fix RMSNorm fp32 input mutation

[JxKim]

Fixes #170.

RMSNorm.rms_forward used an in-place mul_ after x.float(). When the input is already fp32, x.float() returns the original tensor, so the in-place normalization mutates hidden_states and corrupts the residual path.

This changes the scaling step to allocate a normalized tensor instead of mutating the input, and adds a regression test for fp32 inputs.

Testing:

• Attempted: uv run python -m pytest test/test_layernorm.py
• Blocked locally by Windows triton wheel incompatibility, WSL flash-attn build isolation, and later CPU torch dependency download timeout.