ggml-webgpu: move from parameter buffer pool to single buffer with offsets

[reeselevine]

Overview

Continuing some work to simplify and make the WebGPU backend scheduling more asynchronous, I realized that we don't actually need a pool of parameter buffers. Instead we can use a single buffer with multiple offset slots, and cycle through them on a batch of submissions. This PR replaces a pool with a webgpu_param_arena, and moves all operations to use it. Memset is special because it lives in the global context, but because it is now asynchronous it uses a single parameter buffer.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: yes, to help refactor to use the arena and simplify gpu profile future handling

[nikhilJain17]

Nice! This removes a lot of the footguns we were previously dealing with in kernel submission and waiting by getting rid of individual future handles and just waiting on the whole queue.