fix: lmdb optimizations

[Tonguechaude]

So this PR fixes :

1. Deletion of the global Mutex

• Before we have Arc<Mutex<Env<WithoutTls>>> all operation are serialized over only one lock
• Now we have Arc<Env<WithoutTls>> all read operation are concurrent, let LMDB handle his own sync

2. Insert : 2 I/O -> 1 I/O

• Before: get() + put() to write
• Now : get_or_put()

3. Delete : 2 I/O -> 1 I/O

• More or less same as 2. (see delete func)

4. Caching handles (DBI)

• Before: open_database() (mdb_dbi_open called for every operations)
• Now : DashMap to cache handle after the first time

(btw that fix heap crash append cause mdb_dbi_open isnt thread-safe in concurrent)

5. use &str instead of String for table name

• Just because it need to be BlAzInGly fAsT

[Tonguechaude]

I run the benchmark, this PR reduce read operation by 52%

[Tonguechaude]

Single-thread Reads

Size	Before	After	Bonus
512b	412ns	365ns	-9.8%
1kb	415ns	383ns	-7.8%
4kb	593ns	574ns	-6.0%

Concurrents Reads

Threads	Before	After	Bonus
2	6.2µs	6.3µs	no changes
4	13.4µs	12.2µs	-12.8%
8	38.6µs	26.8µs	-35.5%
16	64.3µs	45.6µs	-26.4%