Fix unboxing of numpy boolean scalars in C-backend

[ricardoV94]

We were seeing weird things where an output of ScalarFromTensor for a np.array(True), would evaluate to False. See this old comment here:

pytensor/pytensor/tensor/signal/conv.py

Lines 258 to 260 in d8501d1

	# TODO: Why is .item() needed?
	mode: Literal["full", "valid", "same"] = "full" if full_mode.item() else "valid"
	outputs[0][0] = scipy_convolve(in1, in2, mode=mode, method=self.method)

And this snippet would fail:

import numpy as np
import pytensor
import pytensor.tensor as pt

x = pt.scalar("x", dtype="bool")
y = pt.scalar_from_tensor(x)
fn = pytensor.function([x], y)
assert fn(np.array(True, dtype=bool))   # AssertionError :O

I tracked it down to the c_sync method for ScalarType, which is called when the C code has to generate a Python object as an output (or intermediate computation if going into python mode). Basically unboxing of the internal representation of scalars into numpy scalar arrays.

[Copilot]

Pull request overview

This PR fixes a bug in the C-backend where numpy boolean scalars were not being properly unboxed by the ScalarFromTensor operation. The issue caused np.array(True) to evaluate as False after going through scalar_from_tensor, requiring a workaround using .item() in some places like conv.py.

Key changes:

• Replaced deprecated PyArrayScalar_New and PyArrayScalar_ASSIGN APIs with PyArray_Scalar in the C-backend scalar synchronization code
• Removed the .item() workaround from the convolution operation that is no longer needed
• Added test coverage for boolean scalar conversion to prevent regression

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
pytensor/scalar/basic.py	Refactored c_sync method to use PyArray_Scalar with proper descriptor handling instead of deprecated PyArrayScalar_* APIs, fixing boolean scalar unboxing. Incremented C code cache version.
pytensor/tensor/signal/conv.py	Removed .item() workaround from Convolve2d.perform method that is no longer needed with the C-backend fix.
tests/tensor/test_basic.py	Added test_bool_scalar_from_tensor to verify boolean scalars are correctly converted through scalar_from_tensor without losing their truth value.

[ricardoV94]

Failing test is a precision issue with the new conv2D tests that were merged without a seed, and having too strict tolerance. That's fixed by #811