Edit: See the comments, this construction is not provably secure
After some more thought, the answer is "yes", you can construct a hash from a stream cipher. (My other answer from a few days ago was only negative, eg, how not to construct it.) Here is a construction.
First build a block cipher $B$ from the stream cipher $S$:
Take the PRG $P$ from $S$.
You can build a PRF family $F$ from a PRG using the CGM construction. Use $P$ to build $F$.
You can build a PRP (block cipher) $B$ from a PRF using a Fiestel network. (The Luby-Rackoff theorem states that a secure PRF can be used to construct a secure PRP using 3 (or 4 depending on security desires) rounds of a Fiestel network.) Use $F$ to build $B$.
Now we have a PRP from a PRG. (And we didn't use a hash to do so, as some stream cipher to block cipher constructions do; using a hash in this construction seems uselessly circular). Now build a hash from a block cipher:
You can build a collision resistant compression function $h$ from a PRP using the Davies-Meyer compression function. Use $B$ to build $h$. (Edit: Not necessarily - DM requires an ideal cipher, which is stronger than what $B$ is. The security proof doesn't follow in this step. This may be fixable by finding a different $h$ construction or $B$ construction, but I have not found one.)
You can build a collision resistant hash function $H$ from a collision resistant compression function using the Merkle–Damgård construction. Use $h$ to build $H$.
As to your practical considerations: Is it efficient? I can't speak in general, so let's analyze this construction. Popular hash functions like MD5 and the SHA family use the same Merkle–Damgård construction, so to compare our $M$ against them we'll just focus on analyzing our $h$.
Speed: Every $h$ iteration, we have 3+ rounds of a Fiestel network, each of which consists of keying $F$ and generating a small output from it. Keying our $F$ and generating $n$ bits of output itself requires $n$ re-keys of $P$ and $n$ number of $2n$ bit outputs from $P$ due to how the CGM construction works, and none of it is parrallelizable. So our $n$ bit $h$ requires at least $3n$ re-keys of $P$ and $6n^2$ bits of output from $P$. Considering the speed of modern stream ciphers this is very slow compared to modern hash algorithms, but might be practical in a situation where: hashing is rare, hashed input is small, or hashing is not very time critical.
Implementation overhead: We need to implement the CGM construction, Feistel network, and DM function. Building $F$ from $P$ re-uses $P$'s implementation and adds a feedback loop pushing $n$ bits of output back into the key for $P$ with a branch condition on 1 bit, and no more than $n$ additional bits beyond the input will need to be stored in memory at a time. The Fiestel network construction is minimal overhead, just a few XORs and array swapping. The DM construction is just a couple inputs into the Fiestel network and another XOR. That looks like it is less overhead than a SHA-2 compression function.
We should be able improve on this scheme. We might be able to build a compression function from a PRF more directly than going through the PRP. And the biggest loss of efficiency here is CGM, maybe another construction might turn a PRG into a PRF more efficiently than CGM does.
Edit: This may work if we can build a compression function differently. Until then, it's just an interesting construction with no security proof. I'm leaving it largely as-is as a starting point for a future self or other to fix.