feat: add IO stream infrastructure

[lszskye]

Changes

DataInputStream

• Abstract input stream for reading primitive types (int8, int16, int32, int64, float, double) and byte arrays

DataOutputStream

• Abstract output stream for writing primitive types and byte arrays

ByteArrayInputStream

• DataInputStream backed by an in-memory byte array (Bytes), supporting sequential reads with position tracking

BufferedInputStream

• Buffered wrapper around DataInputStream with configurable buffer size, supporting Seek(), Skip(), and efficient sequential reads

OffsetInputStream

• DataInputStream that reads from a sub-range of an underlying stream, defined by offset and length, with boundary enforcement

MemorySegmentOutputStream

• DataOutputStream backed by MemorySegment with auto-growing capacity, producing MemorySlice output via ToSlice()

Tests

• BufferedInputStreamTest
• ByteArrayInputStreamTest
• DataInputOutputStreamTest
• MemorySegmentOutputStreamTest
• OffsetInputStreamTest

[leaves12138]

I found correctness blockers in the new stream implementations: failed seeks can leave stream positions corrupted, and OffsetInputStream boundary checks truncate 64-bit offsets. Please fix these before merging.

[leaves12138]

Please compute the target position in a local variable and assign position_ only after validation succeeds. Currently a failed seek (for example, seeking past EOF) returns an error but leaves position_ out of range, so subsequent GetPos()/Read() calls observe corrupted stream state. The same invariant should hold for FS_SEEK_CUR and FS_SEEK_END.

[leaves12138]

Same issue here: inner_position_ is mutated before both boundary validation and the wrapped seek succeed. A failed seek should not change the logical position. Please calculate a local target position first, validate it, seek the wrapped stream, and only then commit inner_position_.

[leaves12138]

This boundary check should not take int32_t. The stream API and length_ use 64-bit offsets, so passing values like inner_position_ + size or offset + size through int32_t can truncate offsets above 2 GiB and either reject valid reads or allow out-of-range reads. Please use a 64-bit type consistently here.

[leaves12138]

Re-reviewed PR #39. The remaining concerns about failed seek position rollback and int64_t file system offsets are planned for a later unified file-system refactor, so they should not block this PR. I did not find other blockers.