fix: respect explicit config.head_dim in cnets attention

[AchuthReddy-16]

Fixes #315

Credit to @gcy-hw95, who diagnosed this and tested the fix in #315 — this PR just submits it properly since it never got merged.

This PR fixes the head_dim handling in eagle/model/cnets.py.

Previously, LlamaAttention always computed:

self.head_dim = self.hidden_size // self.num_heads

That works for standard Llama-style configs, but it does not hold for some Qwen3 configs where head_dim is explicitly defined. For example, Qwen3-4B has hidden_size=2560 and num_attention_heads=32, which gives 80 using the old formula, but the actual head_dim is 128.

Because of this mismatch, loading Qwen3 checkpoints can fail with shape mismatch errors.

This PR changes the code to:

• use config.head_dim when it exists
• fall back to hidden_size // num_heads when config.head_dim is not present
• remove the assertion that required hidden_size == num_heads * head_dim
• update the attention output reshape to use num_heads * head_dim instead of assuming it is equal to hidden_size

This follows the same general pattern already used in modeling_qwen3_kv.py.

One note: I left the pretraining_tp > 1 path unchanged because it seems separate from this issue and was not part of the original fix discussed in #315.