perf: encode the Typst source straight into native memory

[msallin]

Encoding.UTF8.GetBytes(inputSource) copied the whole source onto the managed heap, pinned it for one native call, and left it as garbage afterwards. The Rust side copies the bytes into an owned String, so the managed array was never more than a staging buffer.

Measured for a 200 KB source: 200,024 bytes on the managed heap, against zero when encoding into native memory. Anything at or above 85,000 bytes goes to the large object heap, which is only reclaimed by a gen2 collection. This is the path the one-shot statics take, since TypstCompiler.CompilePdf(source) builds a compiler per call.

The change also removes the new byte[1] placeholder. It existed only because fixed over an empty array yields a null pointer, which the native side reads as "no in-memory source"; an explicit one-byte allocation carries that distinction directly. The (pointer, length) contract is unchanged, so a source containing NUL bytes still survives intact.

Also in this change

The native allocations now happen inside the try whose finally releases them. They were made before it, and two statements in between can throw: fontPaths.ToList() over a caller-supplied lazy sequence, and JsonSerializer.Serialize for the system inputs. Either would leak every native block allocated up to that point. Adding one more allocation to that window without closing it would have made an existing leak worse.

Tests

Two safety-net tests added before the refactor and confirmed green against the old implementation:

• EmptySourceIsDistinguishedFromNoSourceAtAll pins the empty-source-versus-null-pointer distinction the placeholder used to carry.
• LargeSourceWithMultiByteCharactersIsCompiledInFull covers a source where UTF-8 byte count and char count diverge.

58/58 tests pass. Builds clean on net8, net9 and net10 with zero warnings.

[msallin]

Reopened upstream against the parent repository: evolvedlight#48