leak global CommandEncoder -- avoid cuda synchronize on process shutdown

[davidkoski]

• used in mlx-swift
• global CommandEncoder are still alive after exit even if no work is happening
• the destructor tries to call cuda api, which causes a fatal error

terminate called after throwing an instance of 'std::runtime_error'
 what(): cudaStreamSynchronize(stream_) failed: driver shutting down

from

frame #11: 0x0000560c45de753f example1`mlx::core::cu::CommandEncoder::synchronize() + 73
frame #12: 0x0000560c45de5c8a example1`mlx::core::cu::CommandEncoder::~CommandEncoder() + 28
frame #13: 0x0000560c45df3ee4 example1`std::pair<int const, mlx::core::cu::CommandEncoder>::~pair() + 32

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: none used

[davidkoski]

Per internal discussion we decided leaking these on process exit was ok. I was able to test this on a cuda device with mlx-swift.

[davidkoski]

I don't know if this has the same issue -- mlx-swift doesn't use the TLS Streams, but since the python code does, perhaps it is ok.