ref(spans): Extract flush_segment pipeline helpers

[lvthanh03]

Refs STREAM-1001

Refactors the flush_segments to be an orchestration function with unit testable helpers:

• _load_flush_candidates: loading ready flush candidates from queue shards
• _acquire_locks_for_flush_candidates: acquiring per-segment flush locks
• _load_segment_data: loading payload keys and payload bytes
• _build_flushed_segments: building producer-ready FlushedSegment objects
• _record_segment_loss_metrics: recording segment loss / expiration metrics

Also adds/moves datamodels (FlushCandidate, FlushedSegment, OutputSpan) into buffer_types.py alongside the existing span buffer value types.

[fpacifici]

One tip to separate observability better:
create a FlushStats class with the following methods:

def add_load_ids_latency
def add_load_data_latency_ms
def add_decompress_latency_ms
def add_segments(segment_keys, segment_to_queue, payloads)

def flush()

Create the object at the beginning of the methods and pass it to the other funcitons.
Each of the functions contributes by adding observability data.
At the end of the method you flush it once.

Now the observability code is totally outside this method.

[fpacifici]

tip:
you are passing two sequences that are supposed to match 1:1. This generally introduces ways this method can break and additional conditions to verify, which increase complexity: what if the two lists do not match ?
This is one of the reasons dealing with the redis api is such a pain.

I'd recommend adjusting the data structures so you have to pass only one (maybe loadedSegmentData) and ensure it contains all the objects you need from flush_candidates.

[linear-code]

STREAM-1001