PERF/mobile: Reduce Nemotron load-time memory peak

[MikeSchirtzinger]

Problem

Nemotron's encoder is materialized as a full JavaScript Uint8Array before ORT session creation. During session build, this likely doubles the largest model allocation and makes mobile/lower-memory browsers much more likely to kill the tab.

Evidence

• nemotron-engine.js:180-190 fetches encoder/decoder/tokenizer bytes, then passes them into WasmNemotron.create(enc, dec, tok).
• nemotron-engine.js:383-400 streams the encoder for progress, but accumulates the full file into new Uint8Array(len).
• crates/nemotron-asr/src/backend_web.rs:159-204 constructs ORT sessions from in-memory bytes.

Acceptance criteria

• Investigate whether ORT web/session creation can avoid a second full encoder copy.
• If possible, stream or transfer bytes into the WASM/worker side to lower peak memory.
• If not possible with current ORT APIs, document the hard blocker and add a memory guard that routes low-memory devices away from Nemotron before loading.
• Measure peak memory before/after or provide a reproducible browser memory trace.