Fix Sequence Header film-grain rewriting with padded AV1 OBUs

[rampageX]

Fix Sequence Header film-grain flag rewriting with padded AV1 OBUs

Summary

This fixes grav1synth apply corrupting some AV1 streams when
film_grain_params_present is not located in the final byte of the declared
Sequence Header OBU payload.

Affected streams can fail with:

obu_forbidden_bit out of range: 1, but must be in [0,0]
Failed to read temporal unit from packet
Error applying bitstream filters to an output packet

Reproduction

One reproducible NVENC AV1 stream used a 3840x1606 Sequence Header.

Before rewriting:

obu_size = 15
film_grain_params_present = 0

After the old write path:

obu_size = 16
film_grain_params_present = 0

The next OBU was then parsed at the wrong byte boundary and its
obu_forbidden_bit became 1.

The clean AV1 stream validates before grav1synth apply, so corruption occurs
during Sequence Header rewriting.

Root cause

There are two related assumptions in the existing write path:

1. sequence.rs uses obu_out.last_mut() when toggling
   film_grain_params_present.

   This assumes the flag is always located in the last byte of the declared
   Sequence Header payload. That is not true when trailing/alignment bytes are
   present.

2. obu.rs adjusts the Sequence Header obu_size by comparing the number of
   bytes written with the number of syntax bytes consumed.

   The syntax parser can stop before trailing/alignment bytes, even though the
   complete declared payload is mirrored into the output. This can falsely make
   the Sequence Header appear to have grown.

Fix

• Compute the real payload-relative byte/bit position of
  film_grain_params_present using TraceCtx::pos().
• Modify that exact byte instead of obu_out.last_mut().
• Treat Sequence Header rewriting as length-preserving and keep its declared
  obu_size unchanged.

Validation

• Existing regression suite passes except for the pre-existing
  film_grain_params_key_frame_forces_update_and_uses_420_zero_luma_shortcut
  test, whose fixture writes grain_scaling_minus_8 = 2 but expects
  scaling_shift == 8 instead of 10.
• The previously failing AV1 stream is the primary reproduction case for this
  fix.

Scope

The change is limited to:

src/parser/sequence.rs
src/parser/obu.rs

No container, encoder, preset, or grain-model behavior is changed.