[js/webgpu] Teach indices helper about int8 and uint8

[hujiajie]

Description

In WebGPU, there is no direct mapping of int8 and uint8 in its supported types. This leads to memory access complexities in the indices helper, depending on how it is configured. The following table summarizes the expected element type of the storage (u32 or atomic<u32>) and how many values each (atomic) u32 encodes (1, 2, or 4).

usage	components = 1	components = 2	components = 4
input	u32, 4	u32, 4	u32, 4
output	atomic u32, 4	atomic u32, 4	u32, 4
internal	u32, 1	u32, 2	u32, 4

Motivation and Context

This is a prerequisite to ease support for int8 and uint8 inside individual operators.

See also the comments for why what configuration takes what path.

[hujiajie]

This is no longer pursued.

#22755 introduces the explicit concept of atomicOutput, which effectively limits the pre-existing output usage to be non-atomic. However, the combo of (u)int8 + non-atomic output + components=1 or 2 cannot be handled in a data-race-free way, as explained in the comments of this PR. It seems quite error-prone to provide a non-atomic helper and ask op implementers to watch out for any data race when calling this.

I also don't love the idea of forcing the components to 4 like bool. The interface just looks poorly designed to me if it consumes a combo of data type + usage + components, while I have to remember a list of (perhaps not so obvious) exceptions.