Numba transpose can be very slow due to ascontiguousarray in dispatch

[ricardoV94]

Description

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

x = pt.matrix("x")
y = x.T

# Avoid deepcopy
x_inp = pytensor.In(x, borrow=True)
y_out = pytensor.Out(y, borrow=True)
c_fn = pytensor.function([x_inp], y_out, mode="FAST_RUN")
numba_fn = pytensor.function([x_inp], y_out, mode="NUMBA")

x_test = np.random.normal(size=(9000, 9000)).T
np.testing.assert_allclose(c_fn(x_test), numba_fn(x_test))
%timeit c_fn(x_test)
%timeit numba_fn(x_test)
# 7.26 μs ± 39.9 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
# 12.8 μs ± 38.5 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)

x_test = x_test[::2, ::2]
np.testing.assert_allclose(c_fn(x_test), numba_fn(x_test))
%timeit c_fn(x_test)
%timeit numba_fn(x_test)
# 7.28 μs ± 190 ns per loop (mean ± std. dev. of 7 runs, 100,000 loops each)
# 108 ms ± 4.87 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

This happens because we (sometimes) introduce a copy here:

pytensor/pytensor/link/numba/dispatch/elemwise.py

Lines 691 to 692 in 19dafe4

	# FIXME: Numba's `array.reshape` only accepts C arrays.
	res_reshape = np.reshape(np.ascontiguousarray(x), new_shape)

We could add a specialized dispatch for pure transpose DimShuffle (without agument or drop dims). Or we could avoid Reshape altogether similar to #847. Although Numba doesn't yet have an overload for np.squeeze and np.expand_dims only accepts one axis at a time. Looking at the implementation it doesn't seem like it would be too hard to extend?