feat(javascript): add bfloat16 and bfloat16_array support

[miantalha45]

Summary

Adds bfloat16 and bfloat16_array support to the Fory JavaScript runtime and codegen, following the same structure as the existing float16 implementation.

Fixes #3288

Changes

Runtime

• BFloat16 class with fromBits() / toBits() and fromFloat32() / toFloat32() (IEEE 754 compliant, round-to-nearest ties-to-even for float32→bfloat16).
• BFloat16Array backed by Uint16Array for packed bfloat16 arrays; supports iteration and fromRaw() for deserialization.
• Wire format: 2 bytes per value (little-endian), matching existing float32/float64 behavior. NaN/Inf/±0/subnormals round-trip correctly.

Type system & serialization

• Type.bfloat16() and Type.bfloat16Array().
• Serializers registered in type resolver; reader returns BFloat16 (no raw uint16 in public API).
• Codegen: gen/builder.ts, gen/number.ts, and gen/typedArray.ts updated for bfloat16 and bfloat16_array.

API

• BFloat16 and BFloat16Array exported from the package.
• Writer accepts BFloat16 | number for scalar bfloat16; arrays accept BFloat16[] or BFloat16Array.

Tests

• Scalar: normal values, NaN, Infinity, ±0 round-trip.
• Array: bfloat16Array with number array and with BFloat16Array.

Checklist

• BFloat16 type with to_bits/from_bits and float32 conversions
• BFloat16Array with packed Uint16Array storage
• Wire format 2 bytes, endianness aligned with float32/float64
• Type.bfloat16() and Type.bfloat16Array(); resolver and codegen
• Tests for scalar and array serialization/round-trip