fix(rtp): keep the voice activity bit out of the audio level's sign

[AlexanderSerbul]

What

When serializing the RFC 6464 ssrc-audio-level extension, the voice activity
(V) bit is forced to 1 for every non-zero level. A header written with
voice_activity: Some(false) is parsed back as Some(true).

The audio level itself is always written correctly — only the flag is lost.

Why

buf[0] = if v2 { 0x80 } else { 0 } | (-(0x7f & v1) as u8);

audio_level is negative, so masking it before negating leaves the sign in
place. For v1 = -37:

step	value
v1 as bits	0xDB
0x7f & v1	0x5B = 128 − 37
-(0x5B) as u8	0xA5 = 128 + 37

Bit 7 is already set before the flag term is OR-ed in, and OR cannot clear it.
The low seven bits come out as the magnitude, which is why the level survived
and only the flag was lost — and why this went unnoticed: with
voice_activity: Some(true) everything matches.

Fix

buf[0] = if v2 { 0x80 } else { 0 } | (v1.unsigned_abs() & 0x7f);

Checked exhaustively over the scale: of the 256 (level, voice_activity) pairs,
127 used to round-trip with the wrong flag — all with the flag false and a
non-zero level — and all 256 round-trip correctly now. The added test
audio_level_round_trip walks the whole scale and fails on main
(left: Some(true), right: Some(false)).

Test changes

The two golden-byte tests in rtp::header encoded the old behaviour: they build
headers with voice_activity: Some(false) and expect 170 (0xAA) where the
correct byte is 42 (0x2A). Updated with a note explaining why, so the change
does not read as a loosened expectation.

The parsing tests are deliberately untouched: 0xAA on the wire really does
mean "voice, level 42", and the parse side needs no change.

cargo test --lib is green (707 passed), cargo fmt --check clean.

How it was found

Building an SFU on str0m. Our own test asserted that a declared "not speech"
flag survives forwarding, and it came back as speech every time. For an SFU this
matters: the point of RFC 6464 is to make forwarding decisions without
decrypting or decoding, and speaker detection that trusts the V bit cannot tell
speech from steady background noise.

[algesten]

Excellent find! Thanks!