feat(parser): emit jsx_text tokens for JSX text in .tsx (#70)

[ericsssan]

Fixes #70 (follow-up to #61/#69, which fixed .jsx only).

Problem

The context-free lexer can't resolve the TSX <T> JSX-vs-generic ambiguity, so #69 gated JSX-text tokenization off for TS. In .tsx, JSX child text stayed lexed as ordinary JS tokens — diverging from espree/@typescript-eslint and from .jsx.

Fix — a single O(n) post-parse pass

The parser already builds correct jsx_text_node/jsx_gap_nodes for .tsx (the AST/ESTree is already right). After parsing, rewriteTsxJsxTextTokens rewrites the token stream to match .jsx:

• Derive edits from the final nodes — each jsx_text_node → collapse its run to one jsx_text token; each jsx_gap_node → insert a jsx_text token for the whitespace gap. Only committed nodes exist (speculative ones were rolled back via nodes.len), so there are no stale edits.
• Two in-place cursor passes — forward to collapse runs (shrink), backward to insert gap tokens (grow, within the lexer's spare capacity; otherwise the gap keeps its token-less, still-ESTree-correct node) — building an old→new token-index map.
• Remap every token-index reference. Audited to be exactly: main_token (all nodes), node_end_toks (all), jsx_text_node.lhs, and the class implements entries in extra_data. (break/continue labels and import/export specifiers store node indices despite misleading ast.zig comments — not remapped.)

In place because the token array is owned by the caller (the Ast only borrows it). Gated to .tsx with JSX text/gaps, so .js/.ts/non-JSX and .jsx carry zero overhead.

Why O(n) (an earlier in-place-during-parse cut was O(n²))

A previous revision collapsed/inserted in place during parsing — one memmove per run → O(n²) (a 4000-element list took ~1.5s). The post-parse pass does it in two linear cursor passes. A 4000-element .tsx list is now single-digit ms.

Result: byte-identical to .jsx

The issue repro <div>\n unrelated{ foo }\n</div> yields JSXText for the content and the trailing gap, matching .jsx exactly. Parity tests assert .tsx == .jsx jsx_text ranges across content, nesting, inter-element + trailing gaps, and generic type-args on a JSX element; TSX generic arrows/calls stay token-free; a remap test extracts import names / labels / implements text after JSX edits (a wrong token index yields wrong text).

Validation

• Full suite green incl. test262 3966/3966 · 1389/1389; TS 17910/17913 · 1210/1223 (the 14 TSX cases feat(lexer): emit one jsx_text token per JSX text child in .jsx (#61) #69 gated off still pass and now carry jsx_text tokens); babel 1928/1928 · 1548/1548; semantic sweep byte-identical to main, 0 crashes.
• Perf: 4000-element list ~single-digit ms (was 1.5s); non-JSX unaffected (the pass is gated to .tsx-with-JSX).
• 3,000,000-iteration ReleaseSafe fuzz of random TSX (3.8M jsx_text tokens): zero crashes, zero non-monotonic token starts, zero out-of-bounds ranges.