Add SINTER, SUNION, SDIFF, SPOP and SRANDMEMBER set commands

[tchalikanti1705]

Summary

Adds the set commands SINTER, SUNION, SDIFF, SPOP, and SRANDMEMBER so RAMen's sets support the multi-set operations and random/pop helpers, alongside the existing SADD/SREM/SMEMBERS/SISMEMBER/SCARD.

Related Issue

Closes #9

Changes

• SINTER <key> [key ...] / SUNION <key> [key ...] / SDIFF <key> [key ...] — set intersection, union, and difference; a missing key is treated as an empty set, and any key holding a non-set returns WRONGTYPE
• SPOP <key> [count] — remove and return one random member as a bulk string (nil if missing), or an array of up to count members; the key is deleted when emptied, and a negative count returns ERR value is out of range, must be positive
• SRANDMEMBER <key> [count] — return random member(s) without modifying the set: no count returns one (bulk, nil if missing), a positive count up to that many distinct members, a negative count exactly -count members with repetition
• Added SInter, SUnion, SDiff (sharing a readSets helper) plus SPop and SRandMember in internal/store/set.go; SRANDMEMBER's with-repetition count is bounded to avoid an unbounded allocation from a client-supplied magnitude
• Added command handlers and registrations in internal/server/cmd_set.go and commands.go
• Added tests in internal/store/store_test.go and internal/server/server_test.go, and documented the commands in docs/commands.md

Checklist

• I have read CONTRIBUTING.md
• Tests pass (go test ./...)
• Code is formatted (go fmt ./...)

[Rohit-Dnath]

Close, and the hard parts are right. One real bug to fix before I merge.

First, the things I specifically tried to break and could not. readSets takes one shard RLock at a time and releases it before the next, so there is no lock-ordering hazard and no same-shard double-RLock, which is the subtle one I was most worried about (RLock twice from one goroutine deadlocks if a writer arrives in between). SINTER a a works. SPOP's bulk-vs-array reply difference between the no-count and count forms is right, the key is deleted when emptied, and SRANDMEMBER's negative count allowing duplicates is right, which is the one people usually get wrong. Expiry uses getLive on the write path and peekLive on the reads. Good work.

The bug: SINTER skips type-checking keys after the first missing one

readSets(keys, stopOnMissing=true) returns on the first missing key, so later keys never get type-checked. The comment says this mirrors Redis. It does not. Redis 7.4's sinterGenericCommand continues past a NULL key, type-checks every key, and only then returns empty:

if (!setobj) { empty += 1; sets[j] = NULL; continue; }
if (checkType(c,setobj,OBJ_SET)) { zfree(sets); return; }

With SADD a 1 and SET str v:

command	this PR	Redis 7.4
SINTER a nope str	[]	WRONGTYPE
SINTER a str nope	WRONGTYPE	WRONGTYPE
SDIFF a nope str	WRONGTYPE	WRONGTYPE

Two things bother me here beyond the Redis mismatch. The result depends on argument order, so the same three keys give an error or an empty array based only on how they were written, which reads as nondeterministic from a client. And SINTER now disagrees with SDIFF and SUNION on identical inputs, which is the kind of inconsistency that costs someone an afternoon later.

The fix drops stopOnMissing entirely: always type-check every key, track whether any was missing, and let SInter return empty if one was. That also collapses readSets back to a single mode and removes the complete return value, so it is less code than what is there now.

Your tests currently assert the wrong behaviour (SInter missing-before-wrongtype in store_test.go, plus the server_test.go case), so those need updating with it. Flagging that explicitly because a passing suite would otherwise make this look fine.

Not blocking

• SRANDMEMBER key -N caps at 2^20 where Redis has no cap. Defensible as a DoS guard and you documented it in the code, but docs/commands.md does not mention it. Worth a line there.
• The multi-key algebra is not atomic, since each shard lock drops before the next. Redis does these atomically. This is the right trade against multi-shard locking and your comment cites the MGET convention, so I am fine with it. Noting it as a known limitation, not a change request.

[tchalikanti1705]

Fixed as you described — dropped stopOnMissing entirely, so every key is type-checked and SINTER agrees with SUNION/SDIFF regardless of argument order. Updated the tests that were asserting the old behavior and added both argument orders over the wire, plus the docs line for the SRANDMEMBER cap. Also rebased onto main.