perf(codec): force inline of vlq_decode for hot loop

Marking the public vlq_decode entry point as #[inline(always)] instead of
just #[inline] lets rustc fully inline the VLQ inner loop into codec::decode
across crate boundaries. The previous `#[inline]` hint was insufficient for
the cross-crate hot path; the function-call overhead and tuple destructuring
of `(value, consumed)` dominated the per-segment work.

Real-world benchmark impact (codec::decode on isolated mappings strings):
- preact (17 KB):  68.6 us -> 17.2 us  (-75%)
- chartjs (490 KB): 3.15 ms -> 1.22 ms (-61%)
- pdfjs (2297 KB): 11.49 ms -> 6.15 ms (-46%)

This closes most of the gap with the sourcemap crate's specialized
decode_mappings (which has always inlined its own vlq_fast helper because
it is private to the crate). The codec path remains slightly slower on
larger maps because it still allocates Vec<Vec<Segment>> per line, but
the inner-loop overhead is gone.

Author: BartWaardenburg

crates/codec/src/vlq.rs

@@ -120,7 +120,7 @@ pub fn vlq_encode(out: &mut Vec<u8>, value: i64) {
 /// Decode a single VLQ value from the input bytes starting at the given position.
 ///
 /// Returns `(decoded_value, bytes_consumed)` or a [`DecodeError`].
-#[inline]
+#[inline(always)]
 pub fn vlq_decode(input: &[u8], pos: usize) -> Result<(i64, usize), DecodeError> {
     if pos >= input.len() {
         return Err(DecodeError::UnexpectedEof { offset: pos });