Validate JSON scalars in place instead of materializing byte lists

[ESRogs]

Important

Stacked on #5 (which sits on roc-lang#10173, which sits on roc-lang#10145). Merge in that order.

Overview

Completes the removal of Str → List(U8) conversions from the JSON parse path. roc-lang#10173 converted whitespace trimming and scalar splitting; #5 converted string scanning and decoding; this PR converts the remaining four scalar/Dec helpers — is_json_number, json_dec_digits_are_zero, trim_json_dec_leading_zeros, and normalize_json_dec_digits. After it, the JSON parse path contains no Str.to_utf8 at all.

Detail

• is_json_number walks the scalar's bytes in place: count_utf8_bytes + str_get_utf8_byte_unsafe replace the to_utf8 List and list_get_unsafe reads; the grammar walk is otherwise unchanged.
• json_dec_digits_are_zero iterates lazily with for byte in Str.iter_utf8(digits).
• trim_json_dec_leading_zeros counts leading zeros in place and takes the remainder with str_drop_first_bytes_unsafe — the dropped prefix is ASCII zeros, so the cut is a proven UTF-8 boundary (previously this revalidated the tail through from_utf8_lossy).
• normalize_json_dec_digits splits the digit string with str_substring_unsafe / str_drop_first_bytes_unsafe at a digit boundary instead of two from_utf8_lossy reconstructions.

Allocation coverage

check_json_parse_allocs! in test/alloc-count/app.roc gains a skipped-numeric-field case on a small (SSO) document. On the base branch it fails with exactly one allocation (is_json_number's to_utf8 copy of the SSO scalar slice); with this PR the whole parse is zero allocations.

Benchmarks

Interleaved against the base branch (#5's tip; median wall time for a loop of 20,000 Json.parse calls decoding { k0 : Str }, default LLVM --opt=speed, macOS arm64):

document	base	this PR
500 numeric fields (499 skipped, validated by is_json_number)	1.16 s	1.13 s (-3%)
500 short string fields (control, no numeric scalars)	0.98 s	0.98 s

The modest delta is expected: to_utf8 on a slice of a heap-backed document was already zero-copy, so the win is the List wrapper and refcount traffic, plus the SSO allocation above.

Testing

JsonScalarParseEdgeCases (51), JsonEncodeNumberEdgeCases (13), JsonStringEscapes (64) all pass; the alloc-count platform app passes on the dev backend and --opt=speed including the new check.

[ESRogs]

Combined with #5 and opened upstream as roc-lang#10217 (branch json-parse-no-byte-lists). Kept for the review history.