refactor(core): Split array.py - separate GPUArray and memory management

[m96-chan]

Problem

src/pygpukit/core/array.py is 741 lines with GPUArray class and memory management mixed.

Current State

src/pygpukit/core/
└── array.py  (741 lines)

Proposed Structure

src/pygpukit/core/
├── array.py             (GPUArray class only, ~400 lines)
├── memory.py            (memory allocation, transfer)
├── dtype.py             (dtype handling, conversions)
└── factory.py           (array creation: zeros, empty, from_numpy)

Benefits

• GPUArray class cleaner
• Memory management isolated
• Easier to swap memory backends
• Better testability

Related

• Core infrastructure refactoring

[m96-chan]

Completed in PR #168