fix(rrm): replace AppendToGlobalBuffer with RecordStagingCopy in UploadBuffer

[JeanPhilippeKernel]

Closes #770

What

UploadBuffer was calling AppendToGlobalBuffer(buf, ...) which routes through RecordGlobalBufferCopy. That function only has SHADER_READ in the barrier access mask — it is designed for mesh geometry that is accessed exclusively as an SSBO via gl_BaseVertex indexing. Generic device-local buffers (bone matrices, particle vertex buffers) need VERTEX_ATTRIBUTE_READ and INDEX_READ as well, which RecordStagingCopy provides.

Fix

Replace the AppendToGlobalBuffer call with the same explicit staging path that UpdateBuffer's device-local branch uses:

• Allocate a host-staging buffer
• Copy CPU data into it via vmaCopyMemoryToAllocation
• Record a vkCmdCopyBuffer + full barrier via RecordStagingCopy
• Track the staging buffer in m_batch_frames for retirement

The standalone destination buffer (buf) is already allocated and stored correctly. Only the upload path changes.

Test plan

• Build passes clean
• Engine runs at 120 fps, no validation errors
• No callers of UploadBuffer exist yet — first real caller is SkinningUploadSystem (Sprint 7)