Three targeted performance fixes in the JS MHTML parser hot path:

1. parse.js: Replace splice(...spread) with push loop for byte
   accumulation. The old `resource.data.splice(len, 0, ...next)`
   spread a Uint8Array into individual arguments on every line of
   MHTML content — O(n) per call in a tight loop over thousands of
   lines. A simple `push(next[i])` loop avoids the spread overhead
   entirely. Truncation via `data.length -= N` replaces splice for
   quoted-printable soft line break removal.

2. util.js decodeBinary: Replace character-by-character string
   concatenation (`data += String.fromCharCode(byte)`) with chunked
   `String.fromCharCode.apply(null, chunk)` joined at the end. The
   old approach was O(n²) due to string immutability; each `+=`
   allocated a new string. Chunks of 8192 bytes stay within the
   call stack limit for `apply`.

3. util.js decodeBase64: Replace `atob(v).split("").map(c =>
   c.charCodeAt(0))` with a pre-allocated Uint8Array filled via a
   direct for-loop. The old approach created two intermediate arrays
   (one from split, one from map) that were immediately discarded.

Benchmarked on synthetic MHTML fixtures (before → after):
  - 10KB:   0.51ms → 0.37ms (1.4x)
  - 1MB:   63.2ms → 43.1ms (1.5x)
  - 10MB:   572ms →  423ms (1.4x)
  - 100MB: 6063ms → 3982ms (1.5x)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>