For overkill discussions about these issues, see
"32-Bit Cyclic Redundancy Codes for Internet Applications" which talks about the kinds of errors that various CRCs are guaranteed to detect: https://users.ece.cmu.edu/~koopman/networks/dsn02/dsn02_koop...
(Proper means one nontrivial cycle e.g. An irreducible generating polynomial of a proper size)
I would be surprised if there's material difference in error detection between sha2 and crc255 (other than the 1 bit difference).
However, crc is useless against an adversary - preimage attacks are trivial. at the same size, it becomes an apple to apple comparison, and whether you choose depends on a speed/preimage requirements.
I'd expect that, no matter how many bits were corrupted, the pattern of the corruption, or the size of the file, a 4-byte blake2 hash would have always a 1 in 4294967296 chance of not catching the error; for crc32, the chance of not catching the error would increase with the size of the file or the number of corrupted bits. On the other hand, a CRC is guaranteed to catch certain corruption patterns (like all burst errors of up to a certain size), while for a cryptographic hash like blake2, there's no such guarantee.
(A cryptographic hash sent together with the data, by itself, is also useless against an adversary, since the adversary can change the hash. You need at least a keyed hash to protect against that.)
A crc32 will miss an error if and only if crc32(error-pattern)=0, which happens in 1:2^32, assuming a full-cycle polynomial. While these patterns are easy to characterize, and are independent of the data (neither is true for a crypto hash), the probabilities are the same.
Is there something that a proper CRC256 would be better at than SHA-256?
The other thing that crc can do that sha256 cannot is an error correcting code of sort; if your errors are limited to a small subset (say, only bursts), you can tabulate the crc of those errors, and then (expected-crc xor computed-crc) is the crc of the error; if it's in your list, you know what the error is. I had actually worked on an device where this was useful some 20 years ago.
That said, it's probably easier to find a reliable implementation of SHA256 than of some CRC64, so, if the added CPU cost doesn't bother you[2], I'd say SHA256 is the best choice.
1 - CRC40 will likely be enough. Those extra 24 bits are a huge safety margin.
2 - And it normally shouldn't, not even for performance sensitive applications. The hash is very fast, and you are doing it before IO. Except for some fast disappearing niches, it just won't matter.
An interesting question about CRCs is whether you can implement them efficiently. CRC32 is possible with some help from the CPU (CLMUL as mentioned, or complete instructions for specific polynomials, like crc32c); eg. on Haswell you get to 14-16 GB/s with a CLMUL approach on CRC32. That's obviously much faster than the hashes I mentioned. With the zlib implementation, on the other hand, you only get about one GB/s and there it's really close to B2b already.
Table-based approaches (classic zlib implementation, or slice-by-n, which needs n-times as many tables) would not work well with longer polynomials; the tables would outgrow L1/L2 cache sizes rather quickly. I'd have to think a bit about the tableless CLMUL approach.
I'm not quite sure why no mainstream x86 extensions have extra stuff for this (VIA PadLock does SHA), but I can take a guess: ARX is already reasonably efficient in software (more so with newer designs, eg. BLAKE2 + AVX2) - unlike AES. And unlike AES there aren't really many problems with side channels, especially when we think about GCM.
>"A CRC32 over a couple megs of data is only good enough to likely catch "gross"
The CRC on TCP MSS or on the ethernet frame would never be higher than 1460 or 1500 byes(or 9K in the case of jumbo frames), how or where would a CRC over multiple megs come into play in regular networking?
CRC is used on jumbo ethernet of of sized of 9K and it works just fine. In fact this is the reason that 9K is the limit on jumbo frames.
My conclusion: A CRC32 over a couple megs of data is only good enough to likely catch "gross" corruption, like truncating the data, not finer corruption like memory and disks tend to do it, especially when we're talking about bit rot.
Also, unless you're using an optimized implementation -- and zlib is not optimized -- using eg. PCLMUL or at least a slice-by-4/8 variant it's a waste of time, because eg. BLAKE2b isn't much slower than zlib CRC, but will catch All The Corruption!1 [1] (Also, Python 3.6 added the full might of BLAKE2 to the stdlib, although it's the reference implementation and not some fancy SSE4/AVX(2) impl.)
[1] A technical note: There is a tradeoff between checksum length and data length; with a good checksum increasing it's length eventually increases the error rate over the same medium (aka false positives [the checksum is corrupted instead of the data]). However, it's usually more important these days to be certain about correctness and not to minimize total indicated BER over a certain channel.