fix: prevent metagraph memory leak (~50MB/20min)

[loayei97]

PROBLEM:

• bt.metagraph() was called on each weight submission cycle (~20 min)
• Each call registers scalecodec types that are never garbage collected
• This caused ~50MB memory leak per iteration, leading to OOM after ~13 hours

SOLUTION:

• Create metagraph ONCE at startup
• Use metagraph.sync(subtensor=...) to refresh data without creating new objects
• Reuse WeightSubmitter instance across iterations instead of creating new

FILES CHANGED:

• run_continuous.py: Reuse WeightSubmitter, call run_submission() instead of run()
• weight_submitter.py: Add run_submission()/cleanup(), use sync() in fetch_metagraph()
• sync_metagraph.py: Create metagraph once in connect_chains(), sync() in getters

MEMORY IMPACT:

• Before: ~150 MB/hour growth, OOM in ~13 hours
• After: Stable memory, no leak