fix: support batched inference in solve_euler by removing hardcoded batch size

[Mr-Neutr0n]

Summary

The solve_euler() method in cosyvoice/flow/flow_matching.py hardcodes the batch dimension to 2 when allocating CFG (Classifier-Free Guidance) tensors. This assumes a single-sample batch (1 conditional + 1 unconditional = 2), which means batched inference with multiple samples silently produces incorrect results — only the first sample's conditioning is used.

What changed:

• Replaced hardcoded 2 with 2 * bsz (where bsz = x.size(0)) in all CFG tensor allocations (x_in, mask_in, mu_in, t_in, spks_in, cond_in)
• Updated the indexing from [0]/[1] to [:bsz]/[bsz:] so that all samples in the batch get their conditioning applied correctly
• The unconditional half ([bsz:]) remains zero-initialized as before, which is the expected behavior for CFG

This fix is backward-compatible: for single-sample inference (bsz=1), the behavior is identical to the original code.

Test plan

• Verify single-sample inference produces identical results to before the change
• Test batched inference with bsz > 1 and confirm all samples receive correct conditioning
• Check that CFG guidance scaling works correctly across the full batch