improvements to scatter / gather

[awni]

• Add contigous flag to indices for scatter / gather
• For scatter: more work per thread when the ratio of indices / output is high
• Add vector copy + possible donation for scatter
• Specialization for contiguous update
• Remove 1D specializations for scatter since they are subsumed by improvements to core kernel

Improves backward of upsampling and mostly closes #1313

Machine	MLX Pre	MLX Post	PT
M1 Max	39.89	4.615	11.93
M3 Max	2.539	1.938	10.77

M1 Max Benchmarks

Benchmark	Pre	Post
ND Index Gather	4.364 (ms)	2.820 (ms)
2D Bench, Factor 0.25	61.164 (ms)	29.541 (ms)
2D Bench, Factor 0.5	48.299 (ms)	16.105 (ms)
2D Bench, Factor 1	45.658 (ms)	13.310 (ms)
2D Bench, Factor 2	44.240 (ms)	10.184 (ms)
2D Bench, Factor 4	44.050 (ms)	9.956 (ms)
2D Bench, Factor 8	44.013 (ms)	10.467 (ms)
2D Bench, Factor 16	44.074 (ms)	11.148 (ms)
2D Bench, Factor 32	44.459 (ms)	12.551 (ms)
2D Bench, Factor 128	115.288 (ms)	47.017 (ms)
1D Bench non-contiguous	34.177 (ms)	2.282 (ms)

M3 Max Benchmarks

Benchmark	Pre	Post
ND Index Gather	2.138 (ms)	1.567 (ms)
2D Bench, Factor 0.25	41.247 (ms)	22.860 (ms)
2D Bench, Factor 0.5	27.324 (ms)	8.087 (ms)
2D Bench, Factor 1	23.801 (ms)	4.532 (ms)
2D Bench, Factor 2	22.912 (ms)	3.309 (ms)
2D Bench, Factor 4	22.696 (ms)	3.022 (ms)
2D Bench, Factor 8	22.665 (ms)	2.974 (ms)
2D Bench, Factor 16	22.661 (ms)	2.984 (ms)
2D Bench, Factor 32	22.656 (ms)	2.987 (ms)
2D Bench, Factor 128	22.631 (ms)	2.985 (ms)
1D Bench non-contiguous	17.672 (ms)	1.329 (ms)