fix(pipeline): preserve original pipeline dtype in from_pipe() by default

[Liauuu]

Description

Fixes #12754

Currently, from_pipe() forces the newly created pipeline (and its shared components) to float32 by default unless torch_dtype is explicitly provided. This causes unexpected VRAM spikes, CUDA OOMs, and introduces a regression in developer experience (DX) when working with float16 or bfloat16 pipelines.

This PR modifies from_pipe() to automatically infer and inherit the source pipeline's dtype if the user leaves torch_dtype=None.

Safety & Edge Case Defense

To perfectly guard against the edge case where shared components might have mixed dtypes:

1. It scans all torch.nn.Module components inside pipeline.components.values().
2. It only inherits and overrides torch_dtype if and only if all components share the exact same dtype (len(dtypes) == 1).
3. If there is a rare mixture of dtypes, it safely leaves torch_dtype = None and skips casting, maintaining the exact original state of the components without breaking any shared weight links.

Testing

• Added a comprehensive mock unit test in tests/pipelines/test_pipeline_utils.py checking that from_pipe correctly inherits float16 and shares the same object reference without unwanted mutations.
• The unit test passes successfully locally:

tests/pipelines/test_pipeline_utils.py::FromPipeDtypeTests::test_from_pipe_preserves_dtype_by_default PASSED [100%]