Let a consumer supply PagedSplats' ranged reads (closes #423)

[xohm]

Why

Every RAD read in PagedSplats goes through fetchRange, which is module-private. Two open needs run into that wall from opposite sides:

• Allow overriding fetchRange to support custom decryption of RAD chunks #423 — serving .rad files with an encryption layer, signed requests, or a custom CDN. The URL is still Spark's; what's needed is to sit in front of the fetch. The reporter is patching dist/spark.module.js by hand on every release.
• No URL at all — a .rad inside a container at an offset, in OPFS or IndexedDB, or a local file the user picked, readable only through File.slice. On a page opened from file:// there is no server to range against either.

requestHeader and withCredentials cover neither case.

What

export interface FetchRangeRequest {
  url?: string;            // rootUrl, or a chunked RAD's sibling file;
                           // undefined when there is no rootUrl
  offset?: number;         // both undefined asks for the whole file
  bytes?: number;
  requestHeader?: Record<string, string>;
  withCredentials?: boolean;
  signal?: AbortSignal;    // the PagedSplats' own, aborted on dispose
}

export type FetchRange = (req: FetchRangeRequest) => Promise<Uint8Array>;

// #423: decorate the fetch of a URL Spark still owns
new PagedSplats({
  rootUrl: "https://example.com/scene.rad",
  fetchRange: async (req) => decrypt(await myFetch(req)),
});

// a source with no URL: a container member, a picked File, OPFS
new PagedSplats({
  fetchRange: ({ offset, bytes }) => container.read(member, offset, bytes),
  fileType: SplatFileType.RAD,
});

• The request carries the requestHeader and withCredentials the PagedSplats was built with, so a replacement can honour them rather than reimplement them.
• Returning fewer bytes than asked for means the file ended; the header probe uses that to stop backing off on a small file. Returning more is not allowed.
• With no rootUrl, fileType must be given — there are no bytes to sniff before the first read — and the error says so.
• RAD only. The rootUrl and fileBytes paths are untouched; nothing changes for anyone who does not pass the hook.

Every RAD read now goes through one routing helper, so the hook covers the header probe, single-file chunks and sibling-file chunks alike. One fix falls out of that: the built-in path now passes the abort signal on the header probe, which it did not before.

Verified

Built from this branch and driven in headless Chromium (SwiftShader) against a 106 MB archive holding a .rad member, with no rootUrl: the hook supplies the header and the chunk reads, the first chunk decodes to 48,196 splats, and opening the file costs 18 KB of reads. The same hook backs a file:// page with no server, through File.slice.