[Feature] Add batch float32 zero-copy angle encoding in qdp-core

[viiccwen]

Reference: #1004

What

Add batch float32 support for angle encoding in QDP core and CUDA kernels.

This includes:

• host-side encode_batch_f32(..., "angle") support
• zero-copy CUDA pointer batch APIs for angle float32 input
• CUDA kernel and FFI support for batched float32 angle encoding
• validation and tests for invalid shapes, null pointers, zero samples, and non-finite inputs

Why

QDP already supports angle float32 zero-copy encoding for single samples, but batch angle workloads still do not have the same float32 core support.

That leaves a gap for the common case where training or inference pipelines already hold angle features as batched CUDA float32 tensors. Adding batch float32 support:

• keeps angle float32 behavior consistent between single-sample and batch paths
• avoids unnecessary dtype promotion for batched angle inputs
• preserves zero-copy integration for framework-owned CUDA tensors
• aligns the angle implementation with the existing amplitude batch float32 pattern