Remove kernel keyword arguments that caused a per-thread copy of prob

[AdityaPandeyCN]

The update kernels had c1/c2 keyword arguments that no caller passes. Keyword arguments make Julia split the kernel into a wrapper and a separate body function, and on GPU that body function was not inlined, so every thread copied prob onto its stack before calling it.

This replaces them with constants (PSO_C1, PSO_C2). It also adds a small inlined rand_static for the SVector random draws, since StaticArrays' rand was not inlined either and used stack for its return value.

Stack per thread from ptxas -v (sm_75, D = 10, default CUDABackend()):

Objective	Kernel	main	this PR	main + always_inline
Rosenbrock	ParallelSyncPSOKernel	336 B	0 B	0 B
Rosenbrock	ParallelPSOKernel	352 B	32 B	32 B
BBOB f8	ParallelSyncPSOKernel	1216 B	1008 B	0 B

For BBOB the remaining stack comes from the objective itself (BBOBFunction holds ~844 B of data and its call is not inlined), so that part still needs always_inline = true.

CPU Core tests pass locally. I haven't run the CUDA tests or timed it on a GPU.