feat: Dedicated parameter logging API (log_param, log_params)

[noahgift]

Summary

MLflow has dedicated log_param() and log_params() APIs. Entrenar uses JSON config but lacks dedicated parameter tracking with versioning.

MLflow Equivalent

• mlflow.log_param(key, value) - Single parameter
• mlflow.log_params(dict) - Batch parameters
• Parameter search and filtering
• Parameter comparison across runs

Current State

Entrenar stores parameters in config: Option<serde_json::Value> on experiments, but:

• No dedicated API for parameter logging per run
• No parameter history/versioning
• No parameter search/filtering
• No type validation

Proposed Implementation

pub struct Parameter {
    pub key: String,
    pub value: ParameterValue,
    pub logged_at: DateTime<Utc>,
}

pub enum ParameterValue {
    String(String),
    Int(i64),
    Float(f64),
    Bool(bool),
    List(Vec<ParameterValue>),
    Dict(HashMap<String, ParameterValue>),
}

impl ExperimentStorage {
    fn log_param(&mut self, run_id: &str, key: &str, value: ParameterValue) -> Result<()>;
    fn log_params(&mut self, run_id: &str, params: HashMap<String, ParameterValue>) -> Result<()>;
    fn get_params(&self, run_id: &str) -> Result<Vec<Parameter>>;
    fn search_runs_by_params(&self, filters: &[ParamFilter]) -> Result<Vec<Run>>;
}

pub struct ParamFilter {
    pub key: String,
    pub op: FilterOp,
    pub value: ParameterValue,
}

pub enum FilterOp {
    Eq, Ne, Gt, Lt, Gte, Lte, Contains, StartsWith,
}

Acceptance Criteria

• log_param() and log_params() methods on storage trait
• Type-safe ParameterValue enum
• Parameter search with filters
• Parameter diff between runs
• CLI: entrenar runs search --param "lr > 0.001"

[noahgift]

Implemented in src/monitor/params.rs - ParamLogger with log_param(), log_params(), get_param(), to_json(), diff() APIs. Full test coverage (32 tests).