Fix/noise floor ratchet, adapt noise to the real floor with recovery in both directions

[usrflo]

Symptom

On long-running nodes (I saw it with an ufo integration branch, which combines feature/repeated-sending-2 + feature/quiet-dwell), direct-packet resends and dwell-gated TX get deferred or suppressed even on a quiet channel.
The radio's reported noise_floor drifts to the -120 clamp and never recovers, so the RSSI-margin LBT checks (isResendChannelActive, isChannelNoisy, isChannelActive with interference_threshold) stay permanently over-sensitive — every send looks like it collides with noise.

Root cause

RadioLibWrapper::loop() calibrated _noise_floor from a 64-sample block, but only accepted samples that satisfied

rssi < _noise_floor + SAMPLING_THRESHOLD   // SAMPLING_THRESHOLD = 14

That filter is a one-way ratchet: it admits ever-lower samples but rejects anything above the current floor, so the block mean can only move down. Over time it walks to the -120 lower clamp and sticks there. The only thing that reset it was resetAGC setting _noise_floor = 0 — but resetAGC is gated on agc_reset_interval, which defaults to 0 (off), and forcing 0 would anyway open a brief permissive LBT window (margin = RSSI − 0) until the next block completes.

The fix

Replace the ratcheted block mean with the median of the 64-sample block:

• Accept every idle sample (!isReceivingPacket()) — no downward bias.
• Sort the 64 samples and take the median (mean of the two middle values). The median rejects transient interference spikes in both directions and recovers upward as well as down.
• Write _noise_floor only after a full block is collected. The previous value stays valid while the next block is sampled — no reset-to-0, hence no permissive LBT window during reconvergence.
• Clamp the result to -120 (lower bound of the radio's RSSI range).
• resetAGC() no longer touches _noise_floor; it only discards the in-progress block (the analog frontend was just reset, so queued samples are stale). _noise_floor itself is left in place because the median estimator no longer needs the hard reset that the ratchet did.

SAMPLING_THRESHOLD is removed (it only fed the ratchet filter).
NUM_NOISE_FLOOR_SAMPLES (64) moves to the header so the sample buffer can be a member array.