bugfix/incorrect lora out dims

[akhoroshev]

I have in-house model for which hidden_size != num_heads * head_size (attention dimension is higher then hidden_size) and lora doesn't work properly. But after this fix, everything works exactly the same as in Python (same generations).

My target modules:

  "target_modules": [
    "v_proj",
    "q_proj",
    "k_proj",
    "down_proj"
  ],

It probably makes sense to make similar changes for kCROSS_ATTN_Q,kCROSS_ATTN_V,kCROSS_ATTN_K

[juney-nvidia]

Close this since it already lands into 0.17.