Use sorted const table + binary search for embedded get()

[SeriousBug]

Summary

Ports the "store file contents statically + binary-search lookup" optimization from upstream rust-embed (commit ce93f7e).

The release/embedded codegen previously emitted a giant match file_path { "a" => ..., "b" => ... } with one arm per embedded file. That is O(n) per lookup and generates a lot of LLVM IR for large asset folders.

This replaces it with a sorted const table looked up via binary search:

• generate_embed_impl now collects (path, embed-tokens) pairs, sorts them by path at macro-expansion time, and emits const ENTRIES: &[(&str, EmbeddedFile)].
• get() does ENTRIES.binary_search_by_key(&path, |e| e.0).ok().map(...) (O(log n)).
• EmbeddedFile::__internal_make is now a const fn so the table can be a const.

The closure/index bindings use unlikely identifiers (__rust_embed_*) because quote!-emitted identifiers in derive macros resolve at the call site and would otherwise collide with user items (the actix example defines a unit struct named index).

Compatibility

• Public API (Asset::get(path)) is unchanged.
• Debug/dynamic mode is untouched.

Testing

• cargo build / cargo test pass.
• cargo test --all-features passes (exercises the embed path via always-embed and compiles the actix example).
• The pre-existing cargo test --release failure in tests/dynamic.rs is unrelated (those tests require debug mode) and reproduces on master.

Benchmarks

Setup: Criterion 0.7 benchmark (benches/embed_lookup.rs) plus a bench_size
example. Both embed a generated fixture folder of N small (~84 byte) files
produced at build time by build.rs behind the non-default bench-fixtures
feature. Runtime numbers use N=2000; binary sizes are reported for N=2000 and
N=5000. Measured locally on an Apple Silicon Mac (release build,
debug_assertions off, so the embedded get() path is exercised). Master
numbers come from a temporary branch combining master's source with the
identical bench code.

Runtime get() (Criterion mean, N=2000)

key position	master	this PR	speedup
first	143.8 ns	47.7 ns	3.0x
middle	357.3 ns	46.8 ns	7.6x
last	155.5 ns	48.5 ns	3.2x
missing	563.0 ns	41.6 ns	13.5x

Binary search is position-independent (~42-49 ns regardless of key), while the
match lookup is O(n)-ish and worst on a middle or missing key. The largest win
is the missing-key case (13.5x).

Stripped binary size (bench_size example)

N	master	this PR	delta
2000	1,217,296 B	1,263,968 B	+46,672 B (+3.83%)
5000	2,653,848 B	2,715,992 B	+62,144 B (+2.34%)

Compile time (recompile of the example crate, N=2000)

	master	this PR
example crate recompile	~7.85 s	~5.41 s (-31%)

Takeaway

The win here is runtime and compile time, not binary size. get() lookup is
3x-13.5x faster and, more importantly, flat with respect to key position and
folder size, removing the linear-scan worst case for misses and middle keys.
The codegen/IR reduction shows up as a ~31% faster recompile of the embedding
crate at N=2000. The final stripped binary is slightly larger (+3.83% at N=2000,
+2.34% at N=5000): the sorted (&str, EmbeddedFile) const table plus the key
array adds a bit of .rodata that the old match did not materialize as
explicitly. If binary size is the priority this is a small regression; if lookup
latency or compile time matters, the trade is clearly favorable.

https://claude.ai/code/session_01MixCkrYqoVEZZbPRRNtL3M

[codecov]

Codecov Report

❌ Patch coverage is 90.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 75.97%. Comparing base (94e4a29) to head (532943e).
⚠️ Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
utils/src/file/embed.rs	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #24      +/-   ##
==========================================
- Coverage   79.46%   75.97%   -3.50%     
==========================================
  Files           9        9              
  Lines         448      566     +118     
==========================================
+ Hits          356      430      +74     
- Misses         92      136      +44     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.