Fix crash counting messages with null content (tool-call logs)

[eeshsaxena]

count_message_tokens (and therefore calculate_prompt_cost) crashes on a normal OpenAI message where content is null:

from tokencost import calculate_prompt_cost
messages = [
    {"role": "user", "content": "What is the weather?"},
    {"role": "assistant", "content": None},  # assistant turn that only carries tool_calls
]
calculate_prompt_cost(messages, "gpt-4o")
# TypeError: expected string or buffer

OpenAI's chat schema makes content nullable, and an assistant message that only carries tool_calls has content: null. The counting loop passes every field value straight into encoding.encode(value), which raises TypeError on None. So any pipeline that estimates cost over real conversation logs with tool calls hits this.

A null field adds no text tokens, so the fix skips None values. Counting a null content now matches counting an empty string. Added a test.