Handle mixed-dtype mismatches in autocast linear and conv wrappers

[JPPhoto]

Summary

Fixes mixed-dtype mismatch failures in autocast-wrapped Linear and Conv2d layers during invocation.

This change:

• fixes CustomLinear so it handles dtype mismatches in the plain path, the bias-only mismatch case, and the sidecar aggregated-parameter patch path
• updates shared patch aggregation to preserve real torch.Tensor params instead of degrading them to meta tensors - adds focused regression tests for plain mixed-dtype inference and sidecar parameter patching, with CPU dtype parametrization for portability

Practical exposure:

• partial-load / autocast execution where activations are float16 or bfloat16 but stored weights or bias remain float32
• bias-only dtype drift after manual casting, state-dict loading, or patch application - sidecar patching during invocation, especially non-LoRA parameter patches that produce residual tensors in the wrong dtype
• mixed-precision inference with adapters or patch systems that introduce tensors at a different dtype than the active execution path
• CPU fallback and non-CUDA environments too, since the underlying failure is a generic PyTorch dtype mismatch (mat1 and mat2 must have the same dtype, self and mat2 must have the same dtype)

In practice, this prevents invocation-time failures that would otherwise appear only under certain precision / patching combinations, making them easy to miss and hard to reproduce.

Related Issues / Discussions

Solves the issue of partial-load execution that I experienced.

QA Instructions

Run the focused mixed-dtype regression tests via pytest tests/backend/model_manager/load/model_cache/torch_module_autocast/custom_modules/test_all_custom_modules.py -k 'mixed_dtype_inference_without_patches or mixed_dtype_sidecar_parameter_patch or bias_only_mismatch'

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[Pfannkuchensack]

Works good.