Fix data race in bss

[talubik]

Fix: BSS data race in multi-core configuration

Hello, I have encountered a data race in the BSS segment when running kernels on multi-core configurations.
All cores start concurrently from _start, and each one was calling memset to zero the entire BSS region.
gridDim and blockDim are BSS globals written by vx_spawn_threads on core 0.
A lagging core's memset would overwrite these values after core 0 had already set them,
causing get_global_size() to return 0 and kernels to produce wrong results or loop indefinitely.

Cause of bug

In kernel/src/vx_start.S, every core executed:

la    a0, _edata
la    a2, _end
sub   a2, a2, a0
li    a1, 0
call  memset

Since all cores start simultaneously, any core can run this memset after core 0 has already
written gridDim/blockDim via vx_spawn_threads, silently zeroing them.

Fix

Remove the BSS memset from _start entirely.

Move BSS zeroing to the host side in vx_upload_kernel_bytes:
vx_copy_to_dev uploads the kernel binary (bin_size bytes), then explicitly
zeros the remaining BSS region (runtime_size - bin_size bytes) before the GPU starts.
This guarantees BSS is cleared exactly once, with no possibility of a race.