docs: show how much memory a larger Ollama context needs

[159753a52]

The context length section ends with "A larger context uses more VRAM and RAM, so size it to what your hardware can hold", but doesn't say how much. This adds a short subsection so people can size num_ctx / OLLAMA_CONTEXT_LENGTH before hitting a CPU offload:

• Ollama reserves the KV cache for the whole context at load time.
• The f16 per-token formula, with one worked row for Llama 3.1 8B (32 layers, 8 KV heads, head dim 128 from its config.json: 128 KiB/token, 4 GiB at 32768, 16 GiB at 131072). Weight size (4.92 GB) is from the llama3.1:8b manifest on registry.ollama.ai.
• OLLAMA_NUM_PARALLEL multiplies it, OLLAMA_KV_CACHE_TYPE shrinks it (linking Ollama's FAQ rather than repeating it), sliding-window / MLA models need less.
• How to confirm with ollama ps (CONTEXT and PROCESSOR columns).

Disclosure: the last sentence links ModelVRAM, a free VRAM calculator that I built (no sign-up, runs in the browser, MIT core at https://github.com/159753a52/llm-vram-calculator). If you'd rather not link third-party tools, I'm happy to drop that sentence; the rest stands on its own.

Checked with npx docusaurus-mdx-checker -c docs/getting-started/quick-start/connect-a-provider (all 9 files compile).