feat(data/json): float support in number parsing

[octalide]

Closes #349.

Problem

std.data.json parsed integer numbers only. Downstream mach-gltf needs full
JSON numeric support (node transforms, accessor bounds, keyframes, material
factors are all floats), so the loader cannot progress past structural parsing
until floats land here.

Change

• Number parsing now accepts the full RFC 8259 grammar (int frac? exp?). A
  number carrying a fraction or exponent becomes a float; an integer-looking
  number keeps yielding an i64, byte-for-byte the same path and error messages
  as before.
• Float values are parsed with correct rounding by reusing
  std.text.parse's bignum decimal→f64 machinery over the zero-copy source
  span — no allocation, no copy.
• Value gains is_float + float_val. New accessors:
  • value_is_float(v) bool — whether the number was parsed as a float.
  • value_float(v) f64 — the float value, or an integer widened to f64, so it
    is defined for any NUMBER.
  • value_number(v) i64 unchanged; documented as the integer-number accessor.
• emit writes floats in shortest round-trippable form (fmt.write_f64).

Enabling change (std.text.parse)

parse_f64 derives its length via str_len, assuming null termination — unsafe
on a span embedded in a larger buffer, which is exactly the zero-copy JSON case.
Factored the body into parse_f64_len(s, len) (bounded by an explicit length);
parse_f64 now delegates. No behaviour change for existing callers.

Tests

json (+14): float with fraction / exponent / signed & uppercase exponent /
negative, integer-stays-integer, mixed int+float array, float in object,
malformed rejection (1., 1e, 1.5e), emit exact bytes, built-float emit,
and parse→emit→parse round-trip. parse (+2): parse_f64_len bounded span (does
not read past len) and zero length. Full suite: 620 passed / 0 failed
(std.data.json 52→66). riscv64 cross smoke test green.