fix: stop fabricating per-core GPU utilisation

[abosnjakovic]

Follow-up to #12, which removed the fabricated per-process GPU column. This removes the same disease one layer down: the per-core GPU display.

The problem

GpuMonitor::get_gpu_info() took the single system-wide residency figure and manufactured one value per core from it:

let core_factor = if i < self.core_count / 2 {
    1.0 + (i as f32 * 0.02)
} else {
    0.9 - ((i - self.core_count / 2) as f32 * 0.02)
};
let core_util = (overall_utilization * core_factor).clamp(0.0, 100.0);

Twenty bars in the header, twenty traces in the Tab → GPU timeline, all derived from one number times a fixed ramp. The comment admitted it ("Apple Silicon doesn't expose per-core GPU stats, so we estimate") while the UI presented the output as measurement.

The change

macOS publishes no per-core GPU breakdown — the gpu_power sampler reports aggregate hardware active residency and nothing else. Deleting the fabrication cascades through the plumbing:

• GpuInfo and GpuCoreInfo are gone. GpuMonitor now exposes utilization() -> f32, a direct read of the atomic the background thread already writes. The history buffer, refresh(), and get_info() were ceremony around that one load.
• detect_gpu_cores() is deleted. Its only consumer was sizing the fake per-core buffers, and the hardcoded table was itself inaccurate for binned SKUs — it mapped every M4 Max to 40 GPU cores, but the 32-core variant ships too.
• DataUpdate::Gpu carries one value instead of a Vec<f32>.
• The gpu G0:12% G1:12% … header line is removed, freeing a layout row. It printed one number once per core, and the KPI header already shows that number once.
• TimelineView::Gpu plots the system-wide trace on its own scale instead of N identical fabricated traces. The Tab cycle keeps all four modes.

Net: 81 insertions, 333 deletions. src/gpu.rs goes from 411 to 277 lines.

Tests

Adds a test for extract_percentage() — the one piece of real parsing logic left in gpu.rs — covering the frequency-residency suffix (12.23% (444 MHz: 12% …)) and a leading core index that must not be mistaken for the value.

Removes test_gpu_detection_expectations and test_cores_to_show_calculation. Both asserted arithmetic on literals declared in the test body, called no production code, and described the per-core GPU layout this PR deletes.

Worth flagging separately: four of the remaining tests have the same problem. test_dot_pattern_generation, test_dot_string_format, and test_core_name_formatting exercise generate_dot_pattern and format_core_line, both of which are defined inside mod tests; test_floating_panel_dimensions recomputes (width / 3).max(35) inline and asserts the result. None of them touch production code. Left alone here since deciding between deleting them and rewriting them against real functions is a separate call.

Verification

All five CI steps pass locally on macOS (rustc 1.96.1): cargo build, cargo test (7 passed), cargo fmt -- --check, cargo clippy --all-targets -- -D warnings, and cargo doc with RUSTDOCFLAGS=-D warnings.

🤖 Generated with Claude Code