Numba rewrites these to integers for specialized implementation

Also need to re-filter warnings because pytestmark.filterwarnings overrides library filters

When default `ignore_updates=True` for inc_subtensor, and boolean indices were rewritten during specialize

Also fix infinity identities for unsigned integers

* Implement simple wrapper Op for `scipy.signal.convolve2d`

* Better shape inference

* conv2d gradient v0

* Implement Convolve2d and gradients

* Add same mode, and FFT support

* Relax test tolerance in float32

* Change seed

* Relax test tolerance in float32

* Add `convolve2d` to `signal.__init__`

---------
Co-authored-by: ricardoV94 <ricardo.vieira1994@gmail.com>

Numba doesn't infer the right type based on the static tuple, but does so with separate boolean variables

Scipy helper doesn't have special handling for ipiv int32 variable, and always assumes it must be cast to a float64

* It did not handle complex values correctly
* It increased compile time with the nested function