Numba: don't materialize index broadcast

[ricardoV94]

Follow up to #2345 that I missed

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

rng = np.random.default_rng(0)
x_val = rng.normal(size=(1000, 8, 5))
i_val = rng.integers(0, 1000, (2500, 1))
j_val = rng.integers(0, 8, (1, 8))
y_val = rng.normal(size=(2500, 8, 5))

x, y = pt.tensor3("x"), pt.tensor3("y")
i, j = pt.lmatrix("i"), pt.lmatrix("j")

take = pytensor.function([x, i, j], x[i, j], mode="NUMBA", trust_input=True)
add_at = pytensor.function([x, y, i, j], x[i, j].inc(y), mode="NUMBA", trust_input=True)
take(x_val, i_val, j_val), add_at(x_val, y_val, i_val, j_val)  # warm up compilation

%timeit take(x_val, i_val, j_val)
%timeit add_at(x_val, y_val, i_val, j_val)

Before:

140 μs ± 8.96 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)
230 μs ± 6.28 μs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

After:

112 μs ± 10.3 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)
143 μs ± 3.54 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

[ricardoV94]

The string codegen for AdvancedSubtensor and AdvancedIncSubtensor is pretty grotesque by now. I suggest we follow up with a rewrite ala vectorize_codegen. Unreadable strings aren't better than unreadable numba intrinsic codegen. The only advantage of this one is you can always dprint the string codegen at the end of the day I guess...