Wrong device specification for host to device in xla runners

[PikaCat-OuO]

#2081 is not fixing the issue.

Example error:
./lc0 benchmark --backend-opts="plugin_path=/kaggle/working/libtpu.so,max_batch=4,steps=1,datatype=bf16,device=1" --minibatch-size=4 --max-prefetch=0

Unhandled exception in worker thread: Buffer passed to Execute() as argument 0 to replica 0 is on device TPU_0(process=0,(0,0,0,0)), but replica is assigned to device TPU_1(process=0,(0,0,0,1)).

This line

lc0/src/neural/backends/xla/xla_runner.cc

Line 201 in c3a160c

devices_[0].get())

should be changed to:

devices_.at(device_).get()