[BugFix] Fix queue checkpoint resume with lazy-initialized scalar buf…

[mohsinm-dev]

Description

• When a queue (OrderedQueue/UnsortedQueue) is created with shape=None, the out buffer starts as
  (max_length, 1). After the first append() with scalar labels (B,), it becomes (max_length,). On
  checkpoint resume a fresh queue still has the (max_length, 1) placeholder, causing load_state_dict to fail
  with a shape mismatch.
• The existing OrderedQueue.load_state_dict override was dead code during Lightning resume because PyTorch's
  recursive loader calls _load_from_state_dict on children, not their load_state_dict override.
• Added _load_from_state_dict to both queue classes and load_state_dict to UnsortedQueue for parity.

Fixes #400