performance improvements

[d-v-b]

running this script: https://github.com/zarr-developers/cast-value.py/blob/main/examples/benchmarks/bench_numpy_vs_rust.py

will compare a python / NumPy implementation of cast_value against the implementation defined in this repo. On my AMD machine, I see numbers like this:

uv run examples/benchmarks/bench_numpy_vs_rust.py
Array size: 1,000,000 elements

Configuration                                    Impl   Throughput     Memory
-----------------------------------------------------------------------------
float64 -> float32 (simple narrowing)           numpy       6.4G/s     3.8 MB
float64 -> float32 (simple narrowing)            rust     972.0M/s     3.8 MB

float64 -> int32 (round nearest-even)           numpy     354.3M/s    16.2 MB
float64 -> int32 (round nearest-even)            rust     141.1M/s     3.8 MB

float64 -> float32 (round towards-zero)         numpy      18.8M/s    72.5 MB
float64 -> float32 (round towards-zero)          rust     119.7M/s     3.8 MB

float64 -> uint8 (clamp, SIMD path)             numpy     318.0M/s    17.2 MB
float64 -> uint8 (clamp, SIMD path)              rust       5.3G/s   976.9 KB

float64 -> int32 (scalar_map: NaN/Inf/-Inf)     numpy     228.9M/s    16.2 MB
float64 -> int32 (scalar_map: NaN/Inf/-Inf)      rust     172.5M/s     3.8 MB

in every case where we are slower than NumPy, we should see if there are optimization opportunities. memory usage looks good across the board though.

[d-v-b]

nerdsnipe @mkitti

[d-v-b]

thread-based parallelism might also be useful

[mkitti]

We seem to be a lot faster than numpy on float64 -> uint8 and float64 -> float32 (round towards-zero).