ciflow/torchtitan/199419: [DCP] Read checkpoint items in parallel in FileSystemReader

What happened

FileSystemReader reads a rank's items one at a time, and torch.load copies each tensor again. Add thread_count (like FileSystemWriter's): items are read on a thread pool, each decoded from its zip header and directory only, and each tensor's bytes are read straight into the target (through a pinned chunk for accelerator targets). thread_count=1 keeps the serial path; the default splits 32 threads across the node's r…

Summary assembled by rule from the sources below

Sources