fix(bench): repair three dead fixture references

[g4bri3lDev]

cargo bench panicked before running a single benchmark. Found while benchmarking dizzy dithering in #68; pre-existing on main.

Criterion constructs every benchmark group before executing any of them, so a single missing fixture takes down the whole run — including the synthetic groups that need no files at all.

Three dead references, not one

reference	reality
katzi.png	exists nowhere in the repo
marienplatz.png	only exists at images/benchmark_only/marienplatz.png
benchmark_only/test7.jpeg	does not exist — so bench_full_res was dead too

The third one means full_res has been non-functional for as long as the fixture was missing, which nobody noticed because the run died earlier anyway.

Fix

• real_images now uses cat.png and olympiapark.png. Both exist, and both are 800×480 like every other file in images/ — which keeps that group's hardcoded Throughput::Elements(800 * 480) honest. I added a comment saying so, because picking a differently-sized fixture there would silently produce wrong throughput numbers rather than an error.
• bench_full_res points at benchmark_only/marienplatz.png (1280×960). Its doc comment claimed 6240×4160, which was wrong for any file present; corrected.
• load_fixture becomes try_load_fixture, warning and skipping instead of panicking. That is the root-cause fix: fixture drift should cost you one group, not the entire benchmark suite.

Verification

cargo bench --bench dithering -- --test now constructs and runs every group:

Testing real_images/burkes_spectra6_auto/night     Success
Testing real_images/burkes_spectra6_auto/vivid     Success
Testing real_images/burkes_spectra6_auto/detail    Success
Testing real_images/burkes_spectra6_auto/daylight  Success
Testing full_res/burkes_spectra6_auto              Success

cargo test --workspace (76 unit + 5 regression suites, 40 fixture comparisons) and cargo clippy --workspace -- -D warnings both clean. No library code touched.