fix(streams): remove quadratic buffering in TextLineStream

[tomas-zijdemans]

Please see: denoland/deno#35881

Why

TextLineStream concatenates its entire pending buffer into every incoming
chunk and rescans it from the start (chars = this.#currentLine + chars).
When a line spans many chunks, each fragment costs O(buffered bytes), making
the total cost O(n²) in the line length. With allowCR: true there is a
second issue: indexOf("\r") runs from position 0 on every loop iteration,
so chunks containing many lines are rescanned once per line.

Real-world impact

Lines shorter than one chunk (~64 KiB for files, ~8–16 KiB for fetch) are
unaffected. But multi-MB single-line records are common — JSONL/NDJSON with
embedded documents or base64 payloads, SSE-style streams — and hit the
quadratic directly. Measured on Deno 2.9.1 (Apple Silicon):

scenario	before	after
8 MiB line, 1460 B chunks	2,262 ms	6.3 ms
8 MiB line, 64 KiB chunks	47 ms	1 ms
64 MiB JSONL file (8 × 8 MiB records), Deno.open → TextDecoderStream → TextLineStream	353 ms	33 ms
16 MiB NDJSON over fetch (4 × 4 MiB lines, localhost)	230 ms	17 ms
common case (100k × 100 B lines)	—	within noise (0.9–1.1x)

There is also a robustness angle: input without newlines forces quadratic
CPU on the consumer, so a stream that never sends a terminator can block the
event loop for seconds.

How

Buffer incoming fragments in an array and join only when a chunk arrives in
which a line terminator can complete; scan with position-based
indexOf(x, start) instead of repeated slicing. Same approach
DelimiterStream in this package already uses. Output is byte-for-byte
identical to the previous implementation (verified with a 3,000-case
differential fuzz across both allowCR modes, random fragmentation
including splits inside \r\n). No API changes.

Added tests: line spanning many chunks (both modes), empty chunks, and
chunk-final \r resolved by the next chunk with allowCR: true.

I used Cursor (Claude) to help investigate and write this change.

---

Note

Recreates #7211, which was closed automatically when I accidentally deleted the head fork. The branch is identical to the original PR head.