Reduce memory usage when reading payloads that span multiple packets.

[bgrainger]

Change ref ArraySegment<byte> to ArraySegmentHolder<byte> so that a mutated ArraySegment<byte> can be passed back through a lambda function.

When run with larger BLOBs (i.e., > 16MiB, the size of one packet), the test results are as follows. ReadBlobsNew allocates 60% less data and becomes 16% faster. Other tests are essentially unchanged.

Before

Method	Mean	StdDev	StdErr	Min	Q1	Median	Q3	Max	Op/s	Gen 0	Gen 1	Gen 2	Allocated
ReadBlobsOldAsync	183,667.8394 us	1,161.7614 us	322.2146 us	181,856.7220 us	182,956.8441 us	183,361.9487 us	184,116.7797 us	185,937.5694 us	5.44	-	-	-	67.01 MB
ReadBlobsOldSync	180,624.3700 us	1,573.1979 us	406.1979 us	178,565.8162 us	179,487.9914 us	180,454.4471 us	181,738.6010 us	184,863.7147 us	5.54	16.6667	16.6667	16.6667	67.01 MB
ReadBlobsNewAsync	202,842.5955 us	1,238.6008 us	319.8054 us	200,103.8980 us	202,173.3350 us	202,988.8227 us	203,793.3685 us	204,539.6862 us	4.93	1275.0000	1275.0000	1275.0000	167.76 MB
ReadBlobsNewSync	200,136.9124 us	675.7203 us	187.4111 us	198,914.1884 us	199,651.0397 us	200,025.8022 us	200,667.8334 us	201,531.4505 us	5	1391.6667	1391.6667	1391.6667	167.67 MB

After

Method	Mean	StdErr	StdDev	Median	Min	Q1	Q3	Max	Op/s	Gen 0	Gen 1	Gen 2	Allocated
ReadBlobsOldAsync	181,221.9569 us	769.7803 us	2,981.3462 us	179,935.0513 us	177,621.6681 us	179,229.9717 us	183,292.2712 us	188,012.1440 us	5.52	-	-	-	67.01 MB
ReadBlobsOldSync	182,072.6833 us	516.8273 us	2,001.6635 us	181,741.0179 us	178,817.1828 us	180,375.7257 us	183,730.9112 us	185,683.7819 us	5.49	-	-	-	67.01 MB
ReadBlobsNewAsync	168,934.5490 us	358.2424 us	1,387.4668 us	168,436.2817 us	167,306.9631 us	167,929.1449 us	169,611.1639 us	172,052.8617 us	5.92	829.1667	829.1667	829.1667	67.12 MB
ReadBlobsNewSync	166,960.6247 us	264.5874 us	989.9953 us	167,157.0712 us	165,231.1244 us	166,155.5539 us	167,565.4565 us	168,570.1445 us	5.99	170.8333	170.8333	170.8333	67.01 MB

[bgrainger]

Since the cache is now owned by the MySqlSession, it could keep alive MySqlConnectionStringBuilder.MaximumPoolSize × maximum payload size bytes.

In general, this shouldn't be a problem because idle sessions will eventually get evicted from the pool (to free client and server resources). However, based on the particular workload it could be possible to get into a scenario where many large buffers are being cached longer than necessary. But it's more likely that the caching will improve performance overall for applications that retrieve a lot of large BLOBs from a database.

Finally, it should be noted that Connector/NET has a similar per-connection buffer (MySqlConnection holds a NativeDriver holds a MySqlStream holds a MySqlPacket holds a MemoryStream) so this is not introducing a new problem specific to this implementation. (It should also be noted that in unscientific local testing I performed, this implementation had 30% fewer bytes in all heaps than Connector/NET when opening a lot of connections, reading a large blob, and keeping all the connections open.)

[bgrainger]

Shipped in 0.19.1.