refactor(cv): read thresholds from runtime tuning + load from NVS on boot

cv_process and helpers (frame_diff, extract_blob, find_centroids) now read
diff_thresh, min_blob_px, max_move, max_missed, and line_offset from
state.tuning instead of file-scope static const constants. The four
thresholds are promoted to file-local constexpr defaults in cv.cpp
(CV_DEFAULT_*) and are no longer part of the public cv.h API — external
code can't depend on them.

cv_process signature drops the line_pct parameter; callers use
state.tuning.line_offset instead. This eliminates the drift hazard of
having two sources of truth (DeviceConfig.line_offset vs
CVTuning.line_offset); the former is deleted.

main.cpp now calls config_load_tuning(g_cv.tuning) after cv_init on boot
so previously persisted tuning survives reboot; logs whether tuning came
from NVS or defaults.

The legacy NVS key "line_offset" is intentionally left alone — harmless
and flash_device.py may still write it during provisioning. Migration is
out of scope.

Tests: 12/12 passing (11 existing + 1 new
test_cv_process_respects_runtime_min_blob proving the runtime-read path).
Flash: 1,414,069 bytes (89.9%).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

firmware/src/config.h

@@ -9,7 +9,6 @@ struct DeviceConfig {
     String hmac_secret;    // 32-byte hex string
     String wifi_ssid;
     String wifi_pass;
-    uint8_t line_offset;   // 0-100, percent of frame height for virtual line
 };
 
 // Load all config from NVS. Returns false if device_id/location_id/hmac_secret missing.