feat(channels): implement custom MarkdownV2 escape with pulldown-cmark (#71)

Replace manual markdown parsing with AST-based event-driven rendering
using pulldown-cmark for robust Telegram MarkdownV2 formatting.

Core implementation:
- TelegramRenderer with state machine for context-aware escaping
- 19 special chars escaped in text, only \ and ` in code blocks
- Bold/italic/strikethrough/header/link/list/blockquote formatting
- UTF-8 safe chunking with char boundary detection and newline preference
- Dedicated escape_url() for link URLs (only ) and \ per spec)

Performance:
- Latency: 2.83µs for 500 chars, 25.73µs for 5KB (3,534x below 10ms target)
- Throughput: 121-970 MiB/s depending on content type
- Single allocation strategy, no excessive copying

Quality:
- 37 tests pass (30 markdown + 5 telegram + 2 memory profiling)
- 99.32% line coverage for markdown.rs
- 0 clippy warnings
- 0 vulnerabilities (cargo deny pass)
- No unsafe code

Dependencies:
- Added pulldown-cmark 0.13.0 (MIT license)
- Added criterion 0.8.0 for benchmarking

Files modified:
- crates/zeph-channels/src/markdown.rs - Complete rewrite with TelegramRenderer
- crates/zeph-channels/src/telegram.rs - UTF-8 safe chunking, removed has_unclosed_code_block()
- crates/zeph-channels/src/lib.rs - Made markdown module public
- crates/zeph-channels/benches/markdown_performance.rs - Comprehensive benchmark suite
- crates/zeph-channels/tests/memory_profile.rs - Memory allocation profiling

Fixes #69

Author: bug-ops

Cargo.lock

@@ -12,9 +12,11 @@ repository = "https://github.com/bug-ops/zeph"
 
 [workspace.dependencies]
 anyhow = "1.0"
+criterion = "0.8"
 eventsource-stream = "0.2"
 futures-core = "0.3"
 ollama-rs = { version = "0.3", default-features = false, features = ["rustls", "stream"] }
+pulldown-cmark = "0.13"
 qdrant-client = { version = "1.16", default-features = false }
 reqwest = { version = "0.13", default-features = false }
 serde = "1.0"