api: allow pipeline draining during flush on short streams

[nschimme]

This is a correctness fix with a minor performance hit.

Refactors faacEncEncode() during end-of-stream flushing to absorb internal look-ahead priming ticks inside the encoder and extends the flush budget for HE-AAC, preventing premature 0-byte returns on short audio streams and truncated SBR tails.

The Problem

1. Short-Stream Tail Loss: The API contract for faacEncEncode() requires callers to pass samplesInput == 0 to flush and poll until 0 bytes are returned. However, the encoder previously applied the look-ahead fill check (frameNum <= LOOKAHEAD_DEPTH) unconditionally. For input streams shorter than the look-ahead depth, initial flush calls hit this gate and returned 0 bytes while still priming. Callers adhering to the API contract interpreted this initial 0 as "drain complete" and stopped calling, discarding all buffered audio tail.
2. HE-AAC Tail Truncation: HE-AAC (HE_V1) utilizes an SBR coded-payload ring (frameFIFO, sized SBR_FRAME_FIFO) that trails the core FIFO by one extra tick. The standard LC flush budget (LOOKAHEAD_DEPTH + 1) returned 0 before trailing SBR envelope payloads were fully emitted, dropping real trailing audio near frame boundaries.

Architectural Fix

Instead of forcing callers to disambiguate "still warming up look-ahead" from "genuinely drained," faacEncEncode() now handles flush ticks via an internal do { ... } while (flushing) loop:

• Flushing Path (flushing == 1): Loops internally to absorb no-output look-ahead priming ticks until a real frame is produced OR the pipeline is fully drained. This guarantees that a 0-byte return strictly signifies a fully drained pipeline.
• HE-AAC Flush Budget Expansion: Expands the maximum flush frame threshold from LOOKAHEAD_DEPTH + 1 to SBR_FRAME_FIFO when HE_V1 is active, allowing the SBR payload ring to drain completely before emitting the final 0-byte return.
• Active Encoding Path (flushing == 0): Retains single-pass execution—returns 0 when accumulating, or proceeds downstream once a full frame is ready.
• Shared Pipeline: Reuses the existing psychoacoustic and bitstream logic across both paths, avoiding code duplication and keeping binary footprint minimal.
• Null-Buffer Safety: Treats inputBuffer == NULL as an implicit flush trigger to guard against caller parameter mismatches.

Testing

Benchmark: https://github.com/nschimme/faac/actions/runs/33096630745

[nschimme]

This fixes the LC side but there is still an open HE lag that I'm chasing. It is safe to merge as it is technically correct.

[nschimme]

Fixed the HE side, this is good to go now