feat(cpu): smoothed RP2350 CPU/Clock telemetry + 0xfc HID report

Add cpu_temp_raw_smoothed() (cmd.cpp/cmd.h): 256-sample average over a
slow EMA (alpha=0.15) so the RP2350 on-die temperature converges to the
true value instead of chasing per-sample 12-bit ADC noise (~1 LSB =
0.47 C). Single source of truth, ADC initialised in one place.

Also export CPU/Clock telemetry on HID feature report 0xfc (set_khz,
cached real_khz, vreg code, ADC ch4 raw). The handler is dormant: live
read over WebHID is not feasible (Chrome requires the report declared
in the HID descriptor, and declaring the OLED Edition vendor reports
breaks DualSense enumeration on Windows). It is harmless and left for a
possible future non-WebHID consumer. Link hardware_adc.

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

src/cmd.h

@@ -11,4 +11,11 @@ bool is_pico_cmd(uint8_t report_id);
 uint16_t pico_cmd_get(uint8_t report_id, uint8_t *buffer,uint16_t reqlen);
 void pico_cmd_set(uint8_t report_id, uint8_t const *buffer,uint16_t bufsize);
 
+// Smoothed RP2350 on-die temperature sensor reading (ADC input 4, 12-bit
+// raw). A single sample is very noisy; this averages a large block and runs
+// a slow EMA so the value converges to the true die temperature instead of
+// chasing per-sample noise. Single source of truth — the OLED CPU screen and
+// the 0xfc web telemetry both call this so device and web always agree.
+uint16_t cpu_temp_raw_smoothed();
+
 #endif //DS5_BRIDGE_CMD_H