[Mosaic GPU] Move barrier allocation to SMEM scratch specs
This is slightly less convenient than our previous approach but it has two main upsides:
1. It lets us automatically emit necessary fences and barriers for use with block clusters
2. It lets us share the same block/cluster barrier for all initializations of mbarriers
This change also moves away from the nvgpu dialect for barriers and allocates them in
dynamic SMEM instead of relying on static SMEM. This should give us more control over
SMEM layouts and alignments, and simplifies the lowering process.
PiperOrigin-RevId: 655493451
