feat: build-time console baud, CSI source filtering, and esp-csi-rs 0.10.0

[VoidReg]

Brings the CLI onto the released esp-csi-rs 0.10.0 and adds the console-throughput and CSI
filtering work that came out of investigating #7.

Depends on csi-rs/esp-csi-rs#29 for the library side; 0.10.0 is already on crates.io, so this
branch builds from a clean clone with no [patch.crates-io] entry — which is why the CLI was
held back until the library shipped.

Console throughput

#7 reports a collector's sample rate collapsing. It is console-bandwidth-bound, not radio-bound —
the radio captures several times what the host receives, and the gap was invisible because the
statistics counters were being cleared by the end-of-run reset before anyone could read them.

The levers, measured on two classic ESP32 boards:

change	effect
baud 115200 → 921600	~7×
serialized log mode	~3×
set-csi --htltf=off (acquisition)	~2.3× (6.5 → 14.7 CSI/s)
defmt instead of println	none (9.26 vs 9.53 CSI/s)

defmt is included because it is the intuitive answer and it is wrong: a CSI line is already
formatted ASCII handed over as a single {=str} argument, so there is nothing left to compress.

• Console baud is now build-time, via ESP_CSI_CLI_UART_BAUD (default 115200, validated
  1..=5_000_000, with cargo:rerun-if-env-changed). Build-time rather than a runtime command:
  changing the baud of the link you are typing on cannot be done from the far end of that link
  without a reconnect, and the value has to be known before the first byte of boot output.
• info declares the baud so a host can confirm what it connected at. CLI_PROTOCOL_VERSION
  stays at 2 — the grammar is unchanged.

CSI filtering — set-csi-filter

Reduces what crosses the console at the source rather than truncating records:

• --peer-mac delivers only frames from one transmitter. Several emitters can share one collector,
  so attribution belongs on the device.
• --min-phy drops frames below a PHY floor, e.g. legacy beacons in an HT-only capture.

Rejected frames count as RX drops, so the loss stays visible in show-stats instead of vanishing.

ESP-NOW modes restored

esp-now-central, esp-now-peripheral, esp-now-fast-collector and esp-now-fast-source sit
alongside the HT20/HT40 emitters. Both families are needed: on chips whose MAC refuses a forced HT
rate, ESP-NOW is the emitter path that works.

Not included, deliberately

An earlier revision added --max-samples to truncate the emitted payload. It was withdrawn after
measurement: set-csi --htltf=off reduces what the radio acquires and beat truncation on both
axes — 23.92 vs 20.98 CSI/s and 506 vs 446 captured. Truncating output to go faster is the wrong
layer when the acquisition itself can be made smaller.

This crate is firmware

Marked publish = false. The build product is an image flashed to an ESP32, so there is nothing
for a registry consumer to depend on; it ships as source plus a git tag, and esp-csi-rs is the
published artifact.

Verification

All five chip targets build against the registry-resolved crate. On two classic ESP32 boards:
22/22 console, statistics, acquisition and filter checks, and 17/17 ESP-NOW and emitter
checks. A 15-config matrix at 115200 and 8 at 921600, with one known pre-existing failure (an
AP/STA run followed by an ESP-NOW run yields nothing on the first ESP-NOW attempt and recovers on
the second — reproduced on a pre-merge image).

Relates to #7.