Shrink TransposeDecoder failure sentinel padding and use inline node storage

Reduces per-chunk setup overhead in `riegeli::TransposeDecoder::Parse`:
1. Shrinks the failure sentinel padding from 255 to
   `kNumFailureSentinelNodes = 64` nodes (the maximum transition step offset is
   `0xff >> 2 == 63`, checked by a `static_assert`).
2. Replaces the heap-allocated `std::vector<StateMachineNode>` with inline
   storage in `Context` (`inline_nodes[128]`: up to 64 active nodes plus 64
   sentinels, ~3 KiB on `Decode`'s stack), falling back to
   `std::unique_ptr<StateMachineNode[]>` for larger schemas, and deletes
   `Context` copy/move because `state_machine_nodes` may point into
   `inline_nodes`.
3. Explicitly initializes `tag_data_size` (and `is_implicit` on sentinels),
   which `std::vector::resize` previously value-initialized.
4. Passes `absl::Span<const StateMachineNode>` (active nodes only) to
   `ContainsImplicitLoop` and replaces its `std::vector<size_t>` with
   `absl::FixedArray<uint32_t, Context::kInlineActiveNodes>`.

PiperOrigin-RevId: 1011381055
3 files changed
tree: 38ad0ef737834d5f87093992fd7f61715062e9c4
  1. .github/
  2. doc/
  3. python/
  4. riegeli/
  5. tf_dependency/
  6. .bazelrc
  7. configure
  8. CONTRIBUTING.md
  9. LICENSE
  10. MANIFEST.in
  11. MODULE.bazel
  12. README.md
README.md

Riegeli

Riegeli/records is a file format for storing a sequence of string records, typically serialized protocol buffers. It supports dense compression, fast decoding, seeking, detection and optional skipping of data corruption, filtering of proto message fields for even faster decoding, and parallel encoding.

See documentation.

Status

Riegeli file format will only change in a backward compatible way (i.e. future readers will understand current files, but current readers might not understand files using future features).

Riegeli C++ API might change in incompatible ways.