Fix multi-byte IME input loss in poll_event

[twokidsCarl]

Problem

When typing CJK characters via IME (Input Method Editor), such as Chinese "你好", only the first character appears in the TextArea. The second character is silently dropped.

Root cause: The terminal delivers multiple UTF-8 characters in a single read() call. poll_event() calls tea_parse_input_with_consumed() once, which parses only the first UTF-8 character and returns consumed = 3 (bytes). The remaining bytes (the second character) are discarded because the consumed value is never used to process the rest of the buffer.

Fix

Add an internal byte buffer (pending_buf / pending_len) to bubbletea_program_t. After parsing, any unconsumed bytes (bytes_avail - consumed) are saved to this buffer. On the next poll_event() call, buffered bytes are consumed before reading from stdin.

This ensures all characters from a single read() are eventually delivered as separate events across multiple poll_event() calls.

Changes

• extension.h: Added pending_buf[256] and pending_len fields to bubbletea_program_t
• program.c: Modified poll_event to check/use buffer before reading stdin, save remaining bytes after parse

Behavior

• Transparent to Ruby layer: poll_event() still returns a single Hash per call
• No API change: The run_loop naturally processes buffered events on subsequent iterations
• Backward compatible: Single-byte ASCII input works identically (buffer stays empty)

Test

Typed "你好" (6 bytes: E4 BD A0 E5 A5 BD) in a Bubbletea TUI with TextArea.

Before: Only "你" appears
After: "你好" appears correctly

Compilation

Verified: compiles cleanly on macOS arm64 with Ruby 4.0.0.