Graphpatch

Adds support for torch 2.5.

What's Changed
* Self-review for 0.2.3 release by evan-lloyd in https://github.com/evan-lloyd/graphpatch/pull/4


**Full Changelog**: https://github.com/evan-lloyd/graphpatch/compare/v0.2.2...v0.2.3

Removes `numpy<2.0` constraint, due to the [latest](https://github.com/huggingface/accelerate/releases/tag/v0.34.0) `accelerate` supporting this version. Note that only `torch>=2.3` are compatible with `numpy 2.*`.

What's Changed
* Self-review for 0.2.2 release by evan-lloyd in https://github.com/evan-lloyd/graphpatch/pull/3


**Full Changelog**: https://github.com/evan-lloyd/graphpatch/compare/v0.2.1...v0.2.2

0.2.1 adds support for `torch` 2.4 and Python 3.12.

* Minor fixes to integration for the latest (as of the time of release) versions of `accelrate`, `transformers`
* Miscellaneous bugfixes
* Extraction configuration option `allow_unused_submodules`. Previously, under "strict" (`error_on_compilation_failure=True`) mode, if a submodule was not invoked during extraction, we considered that an error. This is still the default behavior, but can now be overridden with this flag.

What's Changed
* Self-review for 0.2.1 release by evan-lloyd in https://github.com/evan-lloyd/graphpatch/pull/2


**Full Changelog**: https://github.com/evan-lloyd/graphpatch/compare/v0.2.0...v0.2.1

* Fallback to opaque graph on compilation failure (supports patching input, output, buffers, and parameters)
* Add ExtractionOptions controlling graph extraction
* Better `accelerate` support (now supports CPU and disk offloading)
* `generate()` now supported with patching for `transformers` models implementing `GenerationMixin`
* Support for `torch` 2.2 and 2.3 releases

What's Changed
* Self-review for v.0.2.0 update by evan-lloyd in https://github.com/evan-lloyd/graphpatch/pull/1

**Full Changelog**: https://github.com/evan-lloyd/graphpatch/compare/v0.1.0...v0.2.0

Initial release of `graphpatch`!