fix: use the accelerator API instead of hard-coded cuda/cpu device selection

[li-lizhe]

Fixes #6454

Three call sites still picked the device by hand instead of using ColossalAI's own accelerator abstraction, so on a non-CUDA accelerator they silently fell back to the CPU (fx tracer) or raised (rope cache).

Changes

file	before	after
colossalai/fx/tracer/experimental.py::default_device()	torch.device("cuda:0") if torch.cuda.is_available() else torch.device("cpu")	get_accelerator().get_current_device()
colossalai/fx/passes/meta_info_prop.py::metainfo_trace()	same expression	get_accelerator().get_current_device()
colossalai/inference/utils.py::init_to_get_rotary()	torch.cos(freqs).to(self.dtype).cuda()	torch.cos(freqs).to(self.dtype).to(device), device = get_accelerator().get_current_device()

CUDA behaviour is unchanged: get_current_device() is cuda:<current device> there (the previous expression hard-coded cuda:0), and the rope cache still lands on the current CUDA device.

Verification — Ascend 910B4, torch 2.15.0.dev20260917+cpu + torch_npu, CANN, single process

Before:

torch.cuda.is_available() = False
get_accelerator().name = npu
get_accelerator().get_current_device() = npu:0
--- default_device() ---
default_device() = cpu
--- init_to_get_rotary ---
  File ".../colossalai/inference/utils.py", line 63, in init_to_get_rotary
    self._cos_cached = torch.cos(freqs).to(self.dtype).cuda()
AssertionError: Torch not compiled with CUDA enabled

After (same script, same box, only the three files swapped):

get_accelerator().name = npu
get_accelerator().get_current_device() = npu:0
--- default_device() ---
default_device() = npu:0
--- metainfo_trace ---
source line: device = get_accelerator().get_current_device()
--- init_to_get_rotary ---
cos device = npu:0 | sin device = npu:0
cos dtype = torch.float16 | shape = (65664, 4)

The before/after files were md5-compared with the files in this PR after upload, so the code exercised on the box is byte-identical to the code being submitted.

Not tested

• XPU / MPS backends.
• End-to-end ShardFormer training or inference on NPU — only the three functions above were exercised directly.
• The test container does not have every optional ColossalAI dependency installed (galore_torch, bitsandbytes, diffusers, PIL are absent), so those imports were stubbed while loading the package. No ColossalAI source was modified for the test.