benchmark: add BEAM evaluation

[jinliyl]

Background

BEAM evaluates long-term conversational memory at 128K, 500K, 1M, and 10M-token scales across diverse domains and memory abilities. ReMe needs a scalable BEAM harness to characterize memory quality and resource behavior beyond the context sizes covered by smaller conversational benchmarks.

Changes

• Add a BEAM benchmark workflow under benchmark/beam
• Import the official conversations, questions, annotations, and evaluation assets without committing downloaded or generated data
• Ingest conversations incrementally and preserve turn order, timestamps, domain, scale, and evidence identifiers
• Support representative smoke-test subsets and full 128K, 500K, 1M, and 10M-token tiers
• Reproduce the official evaluation protocol and report results by memory ability, domain, and context scale
• Measure retrieval evidence quality separately from end-to-end answer quality
• Track ingestion time, index size, query latency, model and embedding usage, token usage, and cost
• Make long runs resumable and add configuration documentation, result schemas, and focused fixture-based tests