## Problem

`litellm cost-estimate` (#34686) and `litellm doctor` (#34513) give operators an offline CLI for spend planning and SDK diagnostics, but there is no matching one-shot way to count tokens for a given prompt without writing Python or hitting a provider. The internal `litellm.utils.token_counter()` is the canonical function, but it requires a Python import and a working litellm install to use.

This blocks three common offline workflows:

1. **CI cost gates** — fail the build if a prompt exceeds N tokens for a chosen model. 2. **Model selection** — quickly check whether a long prompt fits a small-context model (8K, 16K) before switching. 3. **Chunking strategy** — measure how many tokens a corpus chapter is, for a fixed tokenizer, before picking a chunk size.

Today, all three require either an inline Python call, a one-off `import litellm; litellm.token_counter(...)` in a REPL, or a round trip to `/v1/messages/count_tokens` (#34243 is a separate proxy endpoint issue, not a CLI).

## Motivation

`litellm.cli` already ships two offline subcommands that follow the same shape (single-purpose, no provider calls, table or JSON output, exit code 0/1/2). A `litellm token-count` s…