BUG: properly handle negative indexes in ufunc_at fast path

[mattip]

Fixes #24147

When avoiding the iter API for faster iteration, negative indexes need to be unwrapped.

Other places this is done in a similar way:

• in mapiter_@name@

  numpy/numpy/core/src/multiarray/lowlevel_strided_loops.c.src

  Lines 1683 to 1684 in c143f7b

  	if (indval < 0) {
  	indval += fancy_dims[i];

  (and more like that)
• in PyArray_MapIterReset

  numpy/numpy/core/src/multiarray/mapping.c

  Lines 2281 to 2282 in c143f7b

  	if (indval < 0) {
  	indval += mit->fancy_dims[i];

• in PyArray_MapIterNext

  numpy/numpy/core/src/multiarray/mapping.c

  Lines 2335 to 2336 in c143f7b

  	if (indval < 0) {
  	indval += mit->fancy_dims[i];

  (and more like that)

Test added that failed before, passes after

[charris]

Thanks Matti.

[seberg]

An bad merge (locally) brought me here, but then I noticed that this can modify the original index array:

assert indxs[0] == -1

would fail. I think it shows that passing pointers or full offsets rather than indices makes more sense here (we need a temporary array anyway), but I have to think about how a hot-fix should look like.
Probably either need a local buffer here, or push the fixing into the indexed loop.

(Maybe local buffer is good, since that is what I would want to do with that hot-fix anyway in some form, even if it may make sense to push the buffer for that case slightly further up.)