missing APIs vs std::simd

[Shnatsel]

Desirable, needs implementation (vectors)

• reduce_sum Implement reduce_sum() #357
• reduce_max/reduce_min Implement reduce_min()/reduce_max() and their precise variants #358

Desirable, needs implementation (masks only)

• Rotations Implement rotate_elements_{left,right} on masks #360
• padded shifts

Not investigated (TODO: move into another category)

• Horizontal bitwise reductions: reduce_or, reduce_xor, etc. Implement reduce_and, reduce_or, reduce_xor #346
• reduce_product Implement reduce_product() #361
• Signed int saturating_abs, saturating_neg, is_positive, is_negative, signum.
• leading_zeros, trailing_zeros, leading_ones, trailing_ones (expensive emulation but maybe still faster than scalar?)
• interleave and deinterleave on masks
• to_degrees, to_radians on floats
• Sign, NaN, infinity, finite, subnormal, and normal classification on floats
• simd_clamp on floats
• Ties-away-from-zero round on floats (round-ties-even is supported; ties away from zero has poor hardware support)

Compatibilty/convenience wrappers

Not sure if we should add all of these, the API surface is large enough as it is.

• to_bits/from_bits on floats (as an API compatibility wrapper around bitcast) Add to_bits/from_bits alias to bitcast() #370
• reverse Add reverse() to both vectors and masks #356
• recip
• first_set/last_set on masks (wrapper around to_bitmask() + scalar popcnt?)
• Boolean-scalar &, |, and ^ in either operand order on masks

Tricky API questions

• extract
• simd_swizzle!
• generic cast between arbitrary same-lane-count vectors (we have explicit to_float/to_int and widen/narrow but not fully generic cast)
• lane-wise abs_diff that always needs to return unsigned values, and std doesn't implement for floats

Not planned

• integer div, rem: no hardware support
• Masked load/store: load_or_default, load_or, load_select (only supported on AVX-512)
• Scatter/gather (impractically slow, often slower than scalar) Add (scalarized) safe gather and scatter ops #171
• Trigonometry, exponents, etc. (likely needs a separate crate with configurable trade-offs, not part of base fearless_simd)
• floating-point remainder: no hardware support, approximations available for special cases, full support complex and expensive (might bundle with trigonometry crate?)