[Fix] road_traffic map_type="2" crashes on GPU reset

[vtalpaert]

Fixes three bugs in vmas/scenarios/road_traffic.py that caused crashes when using map_type="2" on a GPU device, once the initial state buffer has accumulated at least one entry.

Missing return in reset_init_state
When is_use_state_buffer=True, the function set agent position/rotation/velocity from the buffer but is missing a return statement (TypeError: cannot unpack non-iterable NoneType object)
Fixed by adding return ref_path, path_id in the is_use_state_buffer branch.

Device mismatch in reset/record conditions
torch.rand(1) created a CPU tensor while probability_use_recording and probability_record lived on the env device (possibly GPU), raising: RuntimeError: Expected all tensors to be on the same device
Fixed by passing device=self.world.device to both torch.rand(1) in two calls.

Tests added in tests/test_scenarios/test_road_traffic.py:

• test_map_type_runs — sanity check for both map types on CPU
• test_map_type_2_reset_uses_buffer — CPU, reproduces first bug
• test_gpu_map_type_2_rand_device — GPU-only, reproduces device bug