Optimizations for Det should also apply to SlogDet

[ricardoV94]

Description

import pytensor
import pytensor.tensor as pt

x_diag = pt.vector("x_diag")
x = pt.diag(x_diag)
y = pt.log(pt.linalg.det(x))

pytensor.function([x_diag], y).dprint()
# Log [id A] 1
#  └─ Prod{axes=None} [id B] 0
#     └─ x_diag [id C]

_, y = pt.linalg.slogdet(x)
pytensor.function([x_diag], y).dprint(depth=3)
# SLogDet.1 [id A] 4
#  └─ AdvancedSetSubtensor [id B] 3
#     ├─ Alloc [id C] 2
#     ├─ x_diag [id D]
#     ├─ ARange{dtype='int64'} [id E] 1
#     └─ ARange{dtype='int64'} [id E] 1
#        └─ ···

[ricardoV94]

We probably should have linalg.slogdet just return sign(det(x)), log(abs(det(x))) and only later specialize to the SlogDet Op.

Then we don't need to worry about the two forms of Det during linalg rewrites

[ricardoV94]

Another point brought up by @jessegrabowski is that SlogDet doesn't have grad. One more reason to only introduce it as a specialization.