Add ollama filter to strip ANSI spinners and keep only final text output

Problem

When running ollama run <model> "prompt" via SSH or in a terminal, the raw output contains hundreds of ANSI escape sequences (spinners ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏, cursor control [?25l, [?2026h, [K, etc.) before the actual text response.

This is extremely token-wasteful when consumed by an LLM agent — a simple 2-sentence refusal from a model generates ~15KB+ of raw terminal noise vs ~200 bytes of actual content.

Proposed solution

Add an ollama filter (TOML-based or built-in) that:

1. Strips all ANSI escape sequences and spinner characters
2. Keeps only the final generated text output
3. Optionally shows model name and basic stats (tokens/s, duration) in compact form

Example

# Before (raw, ~15KB of ANSI noise)
ollama run phi4:14b "Hello"

# After (rtk filtered, ~50 bytes)
rtk ollama run phi4:14b "Hello"
# → Hello! How can I help you today?

Context

Discovered while running safety audits on local LLMs via SSH — the ANSI overhead from ollama's streaming UI makes LLM-agent consumption impractical without filtering.