Document inplace_on_inputs method and remove TODO comment

[Copilot]

This PR addresses the documentation of the inplace_on_inputs method as requested in the issue. The method allows Op classes to create inplace versions of themselves for specified inputs, enabling memory optimizations while maintaining correctness.

Changes Made

1. Removed TODO comment from pytensor/graph/op.py in the inplace_on_inputs method definition
2. Added comprehensive documentation to doc/extending/inplace.rst explaining the inplace_on_inputs method

Documentation Added

The new documentation section explains:

• Purpose: How the method creates inplace versions of ops for memory optimization
• Current usage: Primarily used with Blockwise operations through PyTensor's rewriting system
• Future extension: Will be extended to support core ops directly
• Implementation guidelines: How to properly implement the method in custom ops
• Best practices: Ensuring only allowed inputs are inplaced and returning appropriate destroy_map

The method signature and expected behavior are clearly documented:

def inplace_on_inputs(self, allowed_inplace_inputs: list[int]) -> "Op":
    """Try to return a version of self that tries to inplace in as many as `allowed_inplace_inputs`."""

This method is automatically used by PyTensor's optimization pipeline and is crucial for the efficient memory management in Blockwise operations. The documentation makes it clear that users typically don't need to call this method directly, but Op developers should understand how to implement it properly.

---

💬 Share your feedback on Copilot coding agent for the chance to win a $200 gift card! Click here to start the survey.