Lock-free stream implementation

[martinling]

This PR adds a new lock-free implementation of the basic underlying data structure required by Packetry: an append-only data stream, backed by a file, for which we require fast random access to any part of the data whilst it continues to grow.

Packetry's current implementation, in the FileVec and HybridIndex types, uses file I/O which adds significant overhead. Both reading and writing require exclusive access to a stream object, which is currently managed by a single Mutex protecting the whole Capture database. This creates contention between the decoder and UI threads.

The new implementation creates two objects for each stream: a unique StreamWriter, and a clonable StreamReader. The writer and any number of readers may operate concurrently without blocking each other.

• StreamReader creates memory-mapped regions on demand to access the file in blocks of a configurable size, which must be a multiple of the system page size. Each reader maintains a fixed-size cache of mappings for its most recently used blocks, using the LruBTreeMap type.

• StreamWriter operates much like a conventional buffered writer, but uses interchangeable write buffers whose lifetimes are managed by Arc. To read stream data not yet written to the file, readers clone the Arc of the current write buffer, which they can then retain access to after the writer moves on to using another buffer.

Atomically switching from one write buffer to another is achieved using the ArcSwap type.

[martinling]

I've added a runtime check that the block size is a multiple of the system page size.

[antoinevg]

👍

[antoinevg]

👍