[JAX] Add F8E4M3FN_F8E4M3FN_F32_X{3,4} DotAlgorithmPreset values
Add `F8E4M3FN_F8E4M3FN_F32_X3` and `F8E4M3FN_F8E4M3FN_F32_X4` to `jax.lax.DotAlgorithmPreset` and `jax_default_matmul_precision` to expose the StableHLO 1.21.0 3 operation and 4 operation `float8_e4m3fn` dot emulation algorithms for `bfloat16` and `float32` inputs, requiring `jaxlib_extension_version >= 514`.
Unlike `BF16_BF16_F32_X3` (where `bfloat16` and `float32` share an 8 bit exponent), `float8_e4m3fn` has a narrower 4 bit exponent, so operands are scaled into the `float8_e4m3fn` range before decomposing each input into an initial `float8_e4m3fn` term and a scaled `float8_e4m3fn` rounding residual. `F8E4M3FN_F8E4M3FN_F32_X4` sums all 4 pairwise `float8_e4m3fn` dot products in `float32` and rescales the result, while `F8E4M3FN_F8E4M3FN_F32_X3` drops the residual by residual product for higher throughput.
PiperOrigin-RevId: 1008855310
