Fix swapped conditions in local_sqrt_sqr rewrite

[WHOIM1205]

Fix: Correct swapped logic in local_sqrt_sqr rewrite

Summary

Fixes a numerical correctness bug in pytensor/tensor/rewriting/math.py
where the rewrite rule local_sqrt_sqr had its conditions swapped.

The previous implementation incorrectly transformed:

• sqrt(sqr(x)) → switch(x >= 0, x, nan) (should be abs(x))
• sqr(sqrt(x)) → abs(x) (should be switch(x >= 0, x, nan))

This caused silent wrong numerical results, especially for negative inputs.

---

Root Cause

prev_op (inner op) and node_op (outer op) checks were reversed:

• The branch matching Sqr(Sqrt(x)) returned abs(x)
• The branch matching Sqrt(Sqr(x)) returned switch(...)

The return values were correct — but attached to the wrong condition.

---

Fix

Swapped the two isinstance conditions so that:

• sqrt(sqr(x)) → abs(x)
• sqr(sqrt(x)) → switch(x >= 0, x, nan)

This is a minimal two-line logical correction.

---

Tests Updated

• Corrected expected graph structure.
• Added numerical tests with negative values.
• Ensured behavior matches NumPy exactly.
• Prevents future silent regressions of this type.

---

Impact

• Restores correct numerical semantics.
• Eliminates silent nan pollution in common patterns like sqrt(x**2).
• Ensures PyMC and downstream users get correct magnitude behavior.

[WHOIM1205]

hey @ricardoV94
Swaps the isinstance checks in local_sqrt_sqr so each rewrite matches the correct expression
Previously
sqrt(sqr(x)) and sqr(sqrt(x)) were transformed into the wrong outputs.
This fixes silent numerical errors and adds proper numerical tests (including negative inputs) to prevent regression

[ricardoV94]

Can you point to the original PR here in the comments? Want to see what might have failed in the review process

[ricardoV94]

I still wouldn't compile and evaluate the functions, the assert_equal_computations shows what the function does already. An independent test would be to compare against the unoptimized function not an expected numerical value

[WHOIM1205]

Thanks for the suggestion. I have updated the tests accordingly. The redundant numerical evaluation and function compilation have been removed and the test now relies on assert_equal_computations for verifying the rewrite behavior.

[WHOIM1205]

hey @ricardoV94 is there anything i can change