Do not raise in linalg Ops

[ricardoV94]

Closes #1078

Instead of adding on_error to every Op, this PR choses the simpler approach, just catch and return nan by default.

This also allows us to simplify the numba implementations quite a lot and get rid of a couple of low-level dispatch for condition number and the like. Perhaps it even speeds up uncached compilation.

Only one Op had this special behavior: Cholesky. For helping with transitioning, on_error="raise" is now implemented symbolically with a FutureWarning. The default switched to on_error="nan", which is the "default" in all lingalg Ops (at least the ones I covered).

This is analogous to us returning 1/0->nan without any error/warning. If it's undefined/can't be computed we return nan, and let the users deal with it. Most of the times the error makes it hader to work with, not easier, as there is no symbolic try/except, while there is symbolic isnan(x).any().

---

I also removed the check_finite argument internally. Users can define it without it having any effect for API compat. The stated reason for this in Scipy is that some implementation of BLAS/Lapack can hang / crash the system with non-finite values. I'm taking a chance here, as those issues may never show up in the versions our users install / modern implementations. We can revisit this later if it proves to be still needed.

I'm marking this as a major breaking change.

---

Finally, I've removed some lapack overloads that were never used anywhere, and inlined others that were used only in one place. I believe it makes code more readable and perhaps uncached numba compilation faster

[jessegrabowski]

Requesting changes because I don't understand why you can just delete all the QR helpers.

[jessegrabowski]

I'm a bit surprised you can just delete all of these. They are used by the other dispatches below. If tests are passing, it might just be that coverage isn't good enough. For example xgeqrf is used by qr_full_no_pivot_impl

[ricardoV94]

You weren't using these overloaded functions, you were going directly to the lapack function (numba_foo, instead of foo or _foo)

[ricardoV94]

You weren't using these overloaded functions, you were going directly to the lapack function from _LAPACK().numba_xgeqrf(dtype)

[ricardoV94]

https://github.com/ricardoV94/pytensor/blob/a4723e67de6cb6c3f068a069dd10dd2fff7ff3ae/pytensor/link/numba/dispatch/linalg/decomposition/qr.py#L319

[jessegrabowski]

yeah my understanding was that I needed to give overloads for those even in that case. I guess not!