tdutils: mmap-backed BufferSlice for large file reads

[Lordron]

During initial sync, DownloadShardState reads the persistent state file
via td::read_file, which allocates the entire file size as one heap
buffer. For current mainnet masterchain state this is ~7 GB held in RSS
across deserialization.

This change adds BufferAllocator::create_reader_mmap(fd, size) that
mmap's the file MAP_PRIVATE and constructs a BufferRaw whose data
pointer points into the mmap'd region. dec_ref_cnt handles munmap on
the last reference. A new td::read_file_mmap() helper wraps this, and
ReadFile::start_up opts into mmap for whole-file reads of 64 MB or
more (offset=0, size=-1).

Existing heap-backed code paths are untouched; BufferRaw adds a single
nullable pointer that defaults to null, and all callers access through
a new BufferRaw::data() accessor that picks between the inline array
and the external pointer.

Measured on a 32 GB host syncing mainnet masterchain state 60660942:
RSS at +5 min uptime: 30 GB -> 9 GB
Swap used at peak: 15 GB -> 65 MB
CPU usage: unchanged
Disk reads: +1 MB/s from mmap page faults (9% util)
DB persistence rate: unchanged

The 64 MB threshold avoids mmap overhead for small files while capturing
the target case of multi-GB persistent state files.